#大模型
200
2026-09-06
okf-memory/okf-agent-memory
S
Git-native persistent memory for AI coding agents. Implements Google OKF v0.2 with sub-300µs in-memory BM25 search, embedded MCP server, and progressive disclosure. Slashes token bloat by 80% with zero external databases or dependencies. Built in pure Go.
2026-10-22
Show HN: Omnara – Open-source alternative to Claude Managed Agents
S
Hey all, Kartik from Omnara ( https://omnara.com) here. We're building an open-source API for running durable agents. You create an agent with an API call, and Omnara runs the agent loop, gives the agent sandboxes to run commands in, and handles secrets, permissions, and RBAC. Each agent's state is stored in Postgres, so agents survive restarts and cost nothing while idle. You can message them and stream their events over the API. You may remember us from our early days, when we made a remote-control phone app for Claude Code and Codex. Since the harness ran on a laptop back then, we had to build a relay to reach it, it stopped when the laptop closed, and moving it to a cloud sandbox meant syncing code and agent state between machines. We didn't think remote-controlling a local agent was end-game for how people would run agents, so we built a new product where the agent runs in the cloud, any client can talk to it through an API, and it can still use your laptop as one of its machines. Anthropic and OpenAI are building hosted versions of this, like Claude Managed Agents and the OpenAI Agents API. Omnara is an open-source version you can run on your own servers and use your own models. It's basically a harness server like the opencode or codex servers, but durable, scalable, with the usual harness features like file and shell tools, subagents, compaction, permission prompts, MCP, skills, memory, yada yada yada. Here's some interesting architectural decisions we made while building this out: Most harnesses, like Claude Code, Codex, OpenCode, Pi, Goose, etc, run the agent loop in one long-lived process that keeps its log in a JSONL or SQLite file. In Omnara, the log is an append-only event log in Postgres, with inputs, model outputs and tool calls stored as typed rows in their own tables. The loop runs on stateless workers in the control plane. A worker runs one iteration of the loop, either a model call or the tool calls it produced, then releases the agent and puts the next iteration on a queue in Postgres. Whichever worker is free picks it up, so each iteration of an agent's loop can run in a different process on a different machine. Each iteration's result commits in one transaction, so the model output, its tool calls, usage and the next wakeup are saved together or not at all. If a worker dies mid-iteration, nothing commits, its lease expires, and another worker retries it. Agents waiting on a user, a tool or a retry don't hold a worker, so one worker can multiplex thousands of agents. Also unlike other harnesses, the machines that the workers (our agent loop) run on aren't where the agent runs its commands. Instead, any macOS or Linux machine with an internet connection, like your laptop or any cloud sandbox, can run a small daemon (omnarad) that connects out to Omnara, so it doesn't need any open inbound ports. The agent then gets tools for that machine, like running a command or editing a file. An agent can have any number of machines connected to it. A fun thing I like to do is I have my laptop and a cloud sandbox connected to the same agent, so when my laptop is on, my agent codes there and I can interact with it, and I have it often push to Github. When I close my laptop, my agent sees the laptop went offline, pulls the latest changes to its cloud sandbox, and continues there. The entire codebase is open source at https://github.com/omnara-ai/omnara, and it’s self hostable, even on a single machine with docker compose. You can also use the hosted version at https://omnara.com. Let me know what you think!
2026-09-03
anthropics/commerce-agents
S
Reference blueprint for building shopping and merchant agents with Claude. Examples in retail, commerce, telecom, and entertainment included.
2026-09-20
Show HN: CUA-S1 – A System One Model for Computer Use
S
Hello HN! We're Dillon and Francesco from Cua. We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their decisions and steps. Some tasks require thinking about a plan, exploring different paths, recovering from failure. Other tasks are a question of making local decisions, like this value should go in this box, or should I check this box, or this element should be ignored. We wondered how far we could go with a small model trained to only make these kinds of decisions. Our inspiration was Typesafe's Jev and its System One Model framing. This is a nod to the dichotomy between thinking quickly, automatically, and intuitively (system 1) vs. thinking slowly, analytically (system 2), as described by Daniel Kahneman. The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice? This kind of model does not generate output token by token like most LLMs do, but rather scores the options you give it, which you can check, trust, and use to drive your app's behavior. CUA-S1 is our answer for narrow, specialized decision models for computer use. Our first release is CUA-S1-FORMS. We built this from ideas and code in jevlike, and then trained a second model just to handle form interactions. It has 706k parameters, and the original checkpoint is 2.8 MB. The first training iteration took less than 30 minutes on synthetic data. Given a set of structured elements and values extracted from a document, it predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields, and does not consider screenshots. Element decisions are scored together, and your code can order the actions, and Cua Driver will execute them one at a time. A first evaluation of this specialist vs. hosted Jev on our form task: - For the whole decision set: 99.7% correct vs 83.6%. - For the subset of steps that require an action: 100% correct vs 96%. - For the subset of steps that are just leaving already-filled fields alone: 100% correct vs 74%. The specialist was trained specifically for this task and convention (just press skip for already filled boxes), while hosted Jev has not been fine-tuned for it, so this is an experiment in scoped specialization. We measured 7-9 ms to score a form locally vs. 260-280 ms per call to hosted Jev including network latency, though those samples measure different things and are not end-to-end form completion times. Our interest here is in the space between a brittle script and a general agent loop. The content and layout of form fields vary enough that scripts get unwieldy, but the set of available decisions can remain narrow and well scoped. We want to explore the possibility of a general agent encountering something novel, and passing well understood decisions over to specialists like this. That is a direction we are looking into. The current release is for forms only. We're open sourced the synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license. Comments welcome! Especially if you are building computer-use agents and have run into a recurring decision that is too variable to script but is too narrow to call another LLM for.
2026-09-27
dzhng/jevgrep
S
Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.
2026-08-30
Show HN: We built open OpenRouter that turns usage into a better model
S
Hi HN, we built an open source model gateway. It's a single place to manage our own self hosted, frontier, and open source models in one place. It’s is rust native, built for concurrency, and implements all the config quirks across models and providers (streaming formats, tool calls, model parameters, rate limits, and different error behavior). The gateway adds under 1 ms for BYOK requests and under 2 ms when Experiential supplies the provider key. It has every major inference provider, and 1000+ models refreshed daily via a codex agent that opens a PR. Compared to other similar projects we’re open source, take no markup, allow you to mix local models with a marketplace, and use your traffic to (opt in) train you a model. Simple routing doesn’t warrant a 10% token markup. The way we do this is given standardized OTel traces, we mine representative real tasks, use text world models to simulate rollouts for various models, apply an LLM judge, and fit a nearest neighbor classifier on top of an embedding of a prompt to decide the optimal model for each request. Usually this can map out a better pareto curve on cost/quality than just calling single models but it’s not perfect. Using these simulations we can also do things like suggesting cache hit optimizations, new model suggestions, and training models. It’s open source, so you can deploy it on your own infrastructure, use our hosted version with 0 markup, or read how we design for maximum availability on our website.
2026-10-01
Show HN: Open-source model routing for coding agents at Astra-level performance
S
A few months ago we started building a model router for coding agents because we thought we could outperform any single model with an ensemble approach. Recently we’ve achieved that milestone and I want to talk about how we did it. First of all, a quick explanation: the Weave Router ( https://github.com/weave-os/router) plugs into any coding agent (e.g. Claude Code or Codex) and intelligently switches between LLMs. So, for example, Astra handles tricky debugging or complex system design tasks, and Deepseek v4 Flash handles simple frontend updates. What we’re announcing today is our new routing model, which we’re calling Weave Router 2.0. We benchmarked 2.0 against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas. On both benchmarks, the router had equivalent pass rates. On Terminal Bench, the router hit 52% of Astra’s cost, and completed tasks 2.2x faster. On SWE Atlas, the router cost 54% as much as Astra and ran 2.5x faster. (Full results on our website at https://weaveos.com/router!) It turns out training a model to route effectively - taking into consideration model capabilities, costs, cache awareness, and more - is a really hard problem! I want to talk about three ways we were able to improve so much over the last few months: 1) a new architecture, 2) larger training data set size, and 3) smarter cache-eviction impact calculation. 1) a new architecture. Our initial approach used an RL model without many priors. While RL is still an important part of the story, the cost of fully exploring the space of routing decisions is very high, so we’ve taken some shortcuts that have significantly improved performance. Consider how large the search space for the routing problem is. Take a typical coding agent session, with ~100 agent turns (i.e. 100 LLM API calls). Technically there are 100 chances to select a model. If we assume a roster of ~10 models (of course there are lots more but we can remove any that are Pareto dominated), then there are 10^100 possible paths through that session. We simply cannot explore all of them! So that's why clever tricks to shrink this space are so important. In particular: we trained a hidden Markov model to trace the session state, then a classifier maps the session to one of a few buckets of similar models. Using the HMM allows us to evaluate not just where a session is currently, but how it got there. We've gotten significantly better performance on bucket selection by incorporating that information - we believe this is because two sessions that might look quite similar to a naive classifier are much better distinguished by this HMM approach. Using this HMM + classifier to select a bucket first significantly shrinks the space to explore, by throwing out most models that could not reasonably serve the given session. This rearchitecture was the single biggest performance unlock! 2) larger training data set size (much less technically interesting but still an important part of the story). By using frontier LLMs to help us label a larger and more diverse set of coding agent sessions, we were able to bootstrap the two models discussed in 1) to a better state, while also providing even richer reward signals for RL. 3) smarter cache-eviction impact calculation. One of the hardest parts of routing well (if you care about saving money) is using the model caches intelligently. We built a subsystem that can calculate the expected value of switching models (and thus paying a high one-time cost to fill up a different cache) much more accurately, helping us avoid costly and unnecessary switches in more cases, while still switching when the benefit outweighs the cost. This is where most of our improvement on cost has come from. We still have a lot of room to continue to improve (we won’t rest until we’re consistently beating Astra/Fable, not just tying!) but matching frontier model performance was a huge milestone for our routing model, and in my opinion validates our initial hypothesis that an ensemble of models can do better than any single model ever could. Our router is open source ( https://github.com/weave-os/router) so anyone can try it out. Or if you prefer you can use our hosted version ( https://weaveos.com/router).
2026-09-19
Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash
S
Hey HN, Henry from Cactus here. We submitted Needle 2 here a few weeks ago, and the feedback in the discussion thread was incredibly valuable, thanks! Thanks to all that feedback, we’ve been able to move quickly to release Needle 3 and I'd love to hear what you think again. The key features: 1) Automation (tool calls & structured JSON output): Needle still doesn't chat by design, its quite challenging to pack general capacity into such small models, so we focus on tool calls and structured JSON. If no tool you declared fits the request, you get an empty list back (note for when playing with the demo). 2) Intelligence Laddering: Every layer (2 to 20) is a deployable subnetwork, so one set of weights, 25 to 121 million parameters at 2-bit, shipping as 8-29MB binaries. On a Raspberry Pi 5 it decodes at up to 4k tokens/sec and prefills at up to 10k. 3) Monarch Hadamard MLP: replaces the dense FFN with three learnable Walsh-Hadamard-initialized Kronecker (Monarch) factor pairs interleaved with per-channel diagonal scales, fixed permutations, a SiLU nonlinearity, and a rank-8 input-conditioned gate, so each token gets a fully mixed nonlinear transform of its d_model channels at O(d√d) parameters and compute instead of the O(d²) a dense 4x-expansion MLP would cost. 4) Performance: On Mobile Actions (phone commands, scored on the exact call) the 20-layer model gets 86.0 through the shipped 2-bit binary; LFM2.5 1.2B is at 82.4, Qwen3.5 0.8B at 76.0, Apple's on-device model at 57.6, all at f16. More results on the link, we do not win everywhere ofc. 5) Multilingual: Needle 3 now supports English, French, Spanish, German, Dutch, Italian, Polish, with more languages coming. 6) Finetuning: You can achieve DeepSeek v4 Flash grade performance on a narrow task with just 4L, stress on "narrow task", we found that production users often prefer tuning before production. 7) Triggers: Grounding is a common challenge for tool call, at least for Needle 2, so we added support case-insensitive regular expressions matched against each request to gate false negatives. 8) Confidence: Every response also carries a calibrated confidence score, the minimum of a judgement on the finished call and its decode probability. Act above your threshold, show the call and ask below it, or escalate to a bigger model. 9) Supported Platforms: macOS, Linux on x86-64, ARM64, ARMv7, RISC-V and MIPS32, Windows x64 and ARM, Android, iOS, watchOS, tvOS, the browser as WebAssembly, and a WASI component. Thanks for reading and as always, thoughts appreciated!
2026-09-11
deepseek-ai/DeepSelect
S
DeepSelect: TopK kernels for DeepSeek Sparse Attention (DSA) and Samplers
2026-09-23
jev-chat/jev-chat-jarvis
S
装在手机上的对话副驾：在微信 / QQ / X / 飞书里读懂对方、给出候选回复、一键填入输入框，发不发由你。非侵入，只读屏幕，不 hook 不改包。
2026-09-17
Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh
S
Hi everybody, we post-trained Qwen 3.8 27B to be more efficient by figuring out which tokens were linked to overthinking and penalizing them without "attacking" the reasoning length directly then fixed the accuracy with a bit of secret sauce (hint On-Policy Distillation) and achieved great results (-58% length, 1.95x speed up, <1% accuracy loss) so we wanted to open-source it and hear the feedback of the community. This is the link to the model: https://huggingface.co/ukisai/Swift-Qwen3.8-27b / It's at 80k downloads in 3 days with independent evals here: https://www.reddit.com/r/LocalLLaMA/comments/1wg7dd5/ukisai_... We are also providing a Free Research Purpose API (OpenAI compatible), courtesy of Nvidia who were kind enough to provide us with the GPUs. It's limited at 5RPM. https://ukisai.com/api/swift/v1/models We also made a GGUF (Q1-Q8): https://huggingface.co/ukisai/Swift-Qwen3.8-27B-GGUF and there's also a few nice community quants with even lower/higher precision (Bartowski: https://huggingface.co/bartowski/ukisai_Swift-Qwen3.8-27b-GG...). The community also created amazing MLX, NVFP4, W4A16 and Uncensored versions you can find on Huggingface. IMPORTANT: Our training approach is not a replacement for the reasoning effort settings, chat templates or token caps but is complementary and targets a completely separate issue (overthinking and "anxiety-like" reasoning loops prior seen in PTQ, but as far as we identified also prominent in BF16 of this size class LLMs as well). Contrary to popular belief, these specific patterns do not contribute to answer quality when properly targeted. (our thesis being: reasoning length IS extremely important and should NOT be shortened by force, but rather optimized). This is also demonstrated bellow in our xhigh vs medium effort benchmark. The goal is to keep xhigh accuracy while reducing only the unnecessary part of thinking. The TLDR of our thought process, research, training and a link to the Meta paper that inspired us is in the first comment. The benchmarks: Swift-27B vs Qwen3.8-27B (BF16, all benchmarks ran x5, thinking effort xhigh) GPQA-Diamond: 88.4% -> 88.3%, 58% fewer median tokens LiveCodeBench v6: 76.8% -> 81.6% (+4.8pp, due to default truncation in LCB it is not performance gain), 46% fewer median thinking tokens Terminal-Bench 2.1: 66.7% -> 65.8%, 39% fewer median tokens MMLU-Pro: 85.5% -> 85.0%, 28% fewer median tokens C-Eval: 90.0% -> 90.6%, 19% fewer median tokens IFBench: 73.5% -> 71.8%, 51% fewer median tokens AIME 2026: 98.7% -> 94.0%, 50% fewer median tokens HMMT (Nov 2025): 99.3% -> 96.0%, 46% fewer median tokens ERQA (vision): 67.5% -> 66.3%, 55% fewer median tokens Token savings hold at every reasoning effort (mean thinking reduction): xhigh 41%, medium 23%, low 26% (albeit with accuracy loses of 1-4% on medium and 1-2% on low which we need further testing for) Swift at xhigh vs the base's own effort settings on GPQA-Diamond (198 questions x 5 seeds): Base xhigh: 88.4%, 6,642 median tokens Swift xhigh: 88.3%, 2,771 median tokens Base medium: 84.1%, 1,753 median tokens So Swift keeps xhigh accuracy at under half the tokens, and beats base-medium by 4pp at roughly 1.6x its tokens. End note: While we are keen on complete open-source, we still need to keep a part of our training and data private, being a new lab. The license is not Apache 2.0, but it only affects companies >$1M. We hope this does not pose a problem for the community and are open to feedback on it. We want to contribute as much as possible to the community and would really appreciate feedback on our work, quantization or Swift model requests. We are working on Swift 3.8 Flash Next and have so far gotten up to -53% thinking token usage. We have strong indicators our methodology is reproducible on other model families as well and are asking the community which ones you want us to optimize next.
2026-08-29
hkqr/my-free-code
S
Open-source multi-provider AI gateway for Claude Code and other coding agents, with model routing, streaming, tools, reasoning, fallbacks, and local model support
2026-09-21
ghuntley/preflight
S
A local proxy that scans LLM requests and attachments for secrets before they reach the model.
2026-09-15
Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost
A
Hey HN, Toby from Nari Labs here. We've been working on making OSS speech models super-fast. Last year, we built Dia, the first OSS text-to-speech model capable of doing natural dialogue. Since then, so many more great speech models have been released to the public. But the market is still dominated by closed source models. We think that's an inference problem. Existing systems such as vLLM / SGLang are not well suited for multimodal inference. To prove this, we built an inference engine specialized for Qwen3-TTS and open-sourced it ( https://github.com/nari-labs/nari-qwen3-tts). Running at sub-50 ms latency at 10 RPS, this showed open models can be run much faster and cheaper. Since then, we've been working hard to bring cheap, fast, and high quality serving to all. And we've even beat closed models at their game! Measured on the highly cited Coval (YC S24) voice AI benchmarks, our Qwen3-TTS endpoint not just is #2 in latency, but #1 in accuracy (WER) compared to 11Labs, Cartesia etc. while being the cheapest endpoint. Our Qwen3-ASR endpoint has the lowest latency and #2 accuracy, just 0.1% away from #1. It is the second cheapest model on the list. It took a lot of clever inference engineering to make these models quick, perform well while keeping costs low. Interestingly, Alibaba's official endpoints seem to perform worse in terms of accuracy and latency compared to ours. But nonetheless, much love to the Qwen team for OSS-ing these amazing speech models. We want to continue to push prices down to make speech technology a commodity - so that every app can have great TTS and STT without worrying about unit costs. We're also working on other parts of audio such as diarization - as well as video and world model inference. More to come!
2026-09-04
sohu-mptc/FlashRec
A
FlashRec is a CUDA-graph engine for generative recommendation: wide beam search (3–5 SID steps, n=50–512+) over a trie-constrained catalog, in-process FP8 serving, and ranked beams on /v1/chat/completions.
2026-08-29
XiaoDuoYa/codex-with-chatgpt
A
ChatGPT thinks. Codex works. Use ChatGPT as the planning brain while keeping the Codex harness.
2026-09-02
Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s
A
I built slotstream, a way to run Qwen3.8-Flash-Next 4-bit on a low-memory mac starting from 16GB, a 125B parameter model that would need 100GB+ memory/RAM, thanks to expert-offloading/ssd-streaming. Easy to install/update, and mac-native using MLX and Swift. It ships with auto-mode, which makes a good tradeoff between memory usage and speed. I'll be implementing and porting the MTP module for speculative decoding next
2026-08-29
JordyZomer/lemmalog
A
A Datalog engine for LLM agent memory: stratified rules, provenance-tracked facts, incremental derivation, and an MCP server that lets your harness use it as a shared brain.
2026-09-19
bespokelabsai/nimble
A
Local typed decisions, contrastive data curation, and model evaluation.
2026-09-12
Hona/openeval
A
Typed prompt-and-judge SDK, CLI, and evidence viewer for agent evaluations
2026-09-23
malevrigns/agent-jev
A
AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated probability distributions back in one ~50ms forward pass. Zero output-token decoding.
2026-10-14
composo-ai/DishwasherBench
S
A dishwasher-loading game on a physical simulation of one real machine, built to benchmark language models and playable in a browser. The game inside is called Spotless.
2026-08-26
Reality-JH/FailoverAI
A
Open-source gateway for reliable image, video and LLM jobs. | 面向可靠图片、视频和 LLM 任务的开源网关。
2026-11-04
shanselman/gputaskman
S
A GPU task manager for Windows, Linux, and WSL: per-process VRAM, GPU activity, and LLM model/KV-cache insights (Windows AMD and NVIDIA).
2026-10-21
Umair/stalemap
S
Tell your coding agent which dependencies changed after its training cutoff. Diffs types and changelogs, writes a short block into AGENTS.md / CLAUDE.md.
2026-10-08
orange2ai/podcast-to-article
S
把多人对谈播客的逐字稿整理成可发布长文的 Agent Skill：换轴成文、人名校准、精简三分之一、去 AI 味自检。An agent skill that turns podcast transcripts into publishable articles.
2026-09-11
xiaYuTian11/maskit
S
数据面具 Data Maskit — 专为大模型打造的本地隐私脱敏网关（请求自动打码，回复流式还原；支持 Cursor / Claude Code / Codex / Pi 等任意可配 Base URL 工具）
2026-10-01
Show HN: Strata – an expressive semantic layer that can say no to your LLM
A
Hello HN, I'm Ajo and I built Strata. I spent 4 years at Netflix solving self service for non-technical business users. I think I cracked it with my unique approach to semantic layer design. The key challenge is balancing expressiveness with ease of use for our non technical colleagues. It just so happens that focus made it work pretty well with LLMs too. Strata is a full stack solution. It includes a semantic layer, dashboards, subscriptions, and google sheets exports. All of it can be done view agent conversation or via MCP. Some key concepts and features: - Names are strict. Only one thing called Revenue can exist in a project. If revenue is mapped to multiple tables, the query grain plus speed will decide table. - Strict naming conventions lends itself to enabling Automatic data blending acrros fact domains. - Partition and Aggregate aware semantic routing. This allows you to load a portion of your data into a faster compute engine like Druid or ClickHose, while keeping full history in trino or something cheaper. The query engine will choose the faster tier if the query can be resolved there. - Complex measures: out of the box support for snapshot, LOD (include/exclude), Compound, and Auto leveling compound measures. - On top of the semantic layer users can create custom calculations that span fact domains. - Another interesting feature is the ability to create cohorts and apply as filter to an entire query or single measure within a query. We are deliberately not building a semantic layer that your BI tools connect to. I think that is a dead end. We have apis, and mcp for anything custom you want to build on top. Try it free on docker: https://strata.do/try/ Docs here to get you going: https://strata.do/docs/ This is an early beta version. It is NOT open source. However, the free tier includes up to 25 users. You should be able to run it entirely on your machine with docker. I would love to hear your feedback. Especially, those working with Looker and Cube. Thanks for reading!
2026-08-27
0xBakeer/qwen38-flash-next-spark
A
Run Qwen3.8-Flash-Next (180B) on a single DGX Spark by keeping its 51B n-gram embedding table on NVMe
2026-08-30
semihtalii/brink
A
Know when you're on the brink — Claude Code, Codex & Cursor usage limits on the edge of your Mac or Windows screen
2026-10-17
Show HN: Gemini 4 Argon plays StarCraft
A
I've refreshed StarSkirmish Bench, a benchmark where LLMs play StarCraft BW by writing code, with results from the flagship models of OpenAI, Anthropic, Google, and Muse
2026-10-07
GGGODLIN/cc-mod-image-view
A
Claude Code mod: thumbnails of pasted images above the prompt instead of bare [Image #n] tags
2026-09-24
yetone/magpie
A
Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.
2026-09-19
Continuum-AI-Corp/OrcaBonsai-27B-Uncensored
A
Runtime behavioral ablation for compressed LLMs. First target: Ternary Bonsai 2 27B — no weight modification or re-quantization. by OrcaRouter team
2026-08-23
dbl8005/sitrep-panel
A
A live local HTML report Claude keeps updated as it works — status board, running narrative, screenshots, newest-first log, live-reload. No cloud, no build step.
2026-10-25
alleyxd/quantpicker
A
Measures which GGUF quant of a model to run on your GPU: real speed and quality loss (KL divergence), not estimates.
2026-10-02
BootLoops-ai/bootloops
A
BootLoops 1.0: certified computational tools and house engines for exact and high-precision physics and quantitative science, built to be driven by LLM agents. MIT; docs CC BY 4.0.
2026-09-26
Badtheorylabs/interference-search
A
Reason over explicit states, many branches at once: merge duplicates, cancel dead ends, advance together.
2026-10-22
ahtx/kindi
A
A custom LLM inference engine for the NVIDIA Quadro K420: Ternary-Bonsai-2-2B at 20+ tokens/s with an OpenAI-compatible server
2026-10-22
xingjian-bai/opal
A
OPAL: a reactive optimizer. A coding agent rewrites the optimizer while your model trains.
2026-10-20
AntoGA/agents-contract
A
AGENTS.md contract for AI coding agents + zero-dependency linter + SessionStart hook. Works with Claude Code, Codex, Cursor, Copilot.
2026-10-17
moritzkremb/claude-eyes
A
Let coding agents watch videos: index once, then skim, search, zoom, and measure editing style. Runs locally.
2026-09-22
Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM
A
Sorry for the pretentious name, I know, I know.. It just contains all the pieces I would like to see a AGI model to have, and I can't stand the temptation. Before throwing rocks at me, please take a glance at the Readme, and I hope it will cover your mood a little bit. So, first of all it does work and you can see the sample from the whole training run here: https://raw.githubusercontent.com/volotat/mini-AGI/refs/head... Here is the scaling law graph I have so far, and it looks very promising: https://github.com/volotat/mini-AGI/blob/main/assets/scaling... The model was built under my deep dissatisfaction so we cannot really train even moderately big models (1B+ scale) on the consumer's hardware. We can inference and fine-tune them for sure, but I would like to have full control over what the model sees over the training run, so it is fully aligned with my interests, not some corporations. I was thinking about for some time and come up with two interesting ideas I thought worth pursuing: MoE with a lot of experts that gets added and pruned from the model while it trains, where only a small subset of of experts are actually in use at any particular moment + batch 1 training on the single continuous stream of data. First allows us to be bounded only by the disk space in terms of number of parameters and load and unload experts only when they are needed. The second (if figured out and it turns out to be doable) allows us to get aways with small VRAM capacity because we do not need to store big randomized batches and their respective gradients. I started brainstorming with Claude and after some time we found an approach that seems to be promising, and low and behold, a few weeks pass and you can see the results yourself. Obviously, I did use AI in the process of making this project and I am pretty sure it would be completely impossible for me to do something like this without it, so I hope it is more than justified. The model is still running over the first of 7.8B characters corpus I selected for training, so the weights are not out yet, and it's about a couple weeks of waiting until they are cooked at the current reading speed. And yeah, the model just read continuous interleaved passages from the dataset, each by 32K characters long each as a single stream. Just as you or I would do. The set up seems to be really simple so you can git clone the project, run it and observe everything for yourself. Thanks for your attention.
2026-09-12
kruzovic7/ai-data-extractor
A
Free open-source extractor for AI coding assistant chat histories. Supports Claude Code, Cursor, Windsurf, Aider, Cline/Roo Code, and more.
2026-09-03
Mercor-Intelligence/ApexAgents-SkyRL-Recipe
A
RL-training knowledge-work agents on APEX-Agents data with SkyRL + Harbor
2026-11-09
ArneshBanerjee/gpt2.s
A
GPT-2 small (124M) in hand-written AArch64 assembly. No libc calls. About 100 tokens/s on an M1.
2026-10-09
0xcro3dile/answerui
A
Open-source alternative to ChatGPT's Intelligent UI. Answers arrive as live charts, forms and calculators. Works with any OpenAI-compatible model, in the cloud or local.
2026-10-09
beianchuqing111/agentic-kb
A
法规/规程知识库智能体:混合检索(bge-m3 稠密+稀疏 → 服务端 RRF → 交叉编码器重排)+ GraphRAG + 带三道闸与审计的 ReAct 智能体。含检索质量评测设施、FastAPI 接口与 React 界面。
2026-10-07
Fanorisky/unstream-proxy
A
Zero-dependency proxy for Anthropic-type upstreams with broken SSE streaming
2026-09-28
MaximeRivest/tiny-classifiers
A
Train a 17M-parameter text classifier in a minute, with labels from a big AI model, and know when it beats an API. Recipe, tutorial, results on 10 tasks, phone trainer.
2026-08-31
alesha-pro/atlas
A
Interactive canvas for taking an LLM apart tensor by tensor: measured INT8/INT4/FP8 error, distributions, spectra and outlier channels for every weight tensor
2026-08-23
DavidCarliez/cover
A
Reversible privacy proxy for AI agents: send realistic fakes, restore originals locally.
2026-10-24
Xinge2333/codex-quota-bar
A
macOS Codex quota menu bar with an optional desktop overlay. 中文 / English guides.
2026-10-21
dominikmartn/harp
B
Run many Claude Code and Codex agents side by side, in one sidebar that tells you which one needs you. An unofficial fork of Warp.
2026-10-16
SakanaAI/AutoIllustrate
A
Generate publication-ready, editable draw.io figures for research papers from a methods section, using LLM agents.
2026-10-01
Show HN: Token compression CLI to save Codex/Astra costs
B
Hey HN! Yolanda and Spencer here - wanted to share a token compression tool that we’ve built for ourselves to save 30% costs on codex! After maxing out sub and burning $700/day per person on api, we fine tuned a compression model to trim codex's tool call output to reduce input token + cache. It cut down tokens by 29.6% and now I just leave it on by default in Codex. To avoid messing up w/ cache, we use proxy + fine tuned qwen model trained on preserving agent trajectory to remove tool call results before they go back to the model, leaving kv cache untouched. The cli is free for everyone to use ( https://github.com/spenmcke/compress). Just lmk ur feedback and hacks to shave even more costs on astra! If you want to integrate it into your product to offer the best models at low cost, I can set you up with an sdk and api keys PS: It’s built for coding agents, not conversational agents. I optimized it for file retrieval accuracy, trajectory preservation, and quality to get up to 30% cost reduction depending on how context-heavy the task is. On security and privacy side, it's a proxy wrapping your local codex and ZDR so it doesn't retain any queries. It’s on by default in codex and when you don’t want compression, you can use `codex --uncompress` to disable it. Give it a try: code is in https://github.com/spenmcke/compress You can install the cli using `curl -fsSL https://install.everestagi.com/install.sh | sh && source ~/.config/everest/shell.sh` Love to hear any feedback and learn your hacky ways to save token costs too!
2026-09-21
arikchakma/opencloak
A
Swap the personal details out of your AI prompts before they are sent. On-device detection, no server, no account.
2026-09-20
logan-markewich/jeff
B
A self-hosted drop-in replacement for TypeSafe's jev, powered by GliFormer.
2026-09-17
Krimchanin/claude-vault
A
Keep Claude Desktop chats safe when switching accounts — local-first backup and restore for conversations.
2026-08-28
tonyd2wild/GLM-5.3-Flash-NVFP4-DFlash2-2x-DGX-Spark
B
GLM-5.3-Flash (NVFP4) on 2x NVIDIA DGX Spark - vLLM TP2, 262K context, MTP. World-first deploy recipe: 7 day-0 bugs found and fixed, patched sm121 image, probes and full report.
2026-11-03
Balanov/verdict
A
A 197 KB local semantic verifier for AI agent tool calls. Exit code 0 is not proof.
2026-11-02
iiinno/OpenHinoVoice
A
Windows で使う日本語の音声入力アプリ。認識はローカル、整形は ChatGPT / OpenCode Go / OpenAI から選べる
2026-10-13
Dushiqi1987/codex-remote-bridge
A
本地 Codex 连接服务器 Codex CLI：通过 SSH 反向隧道共享本地 HTTP 代理或直连网络，适用于 AutoDL 等 GPU 科研服务器。含 Python 转发器、PowerShell 隧道脚本、中文教程与测试。
2026-10-11
liyupi/ai-promo-video-maker
A
一句话让 AI 帮你做产品宣传片！把产品名或网址告诉 Cursor、Claude Code、Codex 等 AI 编程工具，它会自己调研产品、画出多种风格样图让你挑选、截取 4K 真实页面、按音乐节拍卡点剪辑、生成 AI 旁白并逐句转写校验、自动审图，最后交付一支 4K 商业宣传视频和可在线播放的网页版。App、网站、SaaS、开源项目、课程、AI 模型都能做，你只需确认需求、试听配乐。不用学剪辑软件，也不用花钱找外包。基于 Remotion + ElevenLabs，可直接克隆使用，也能一键安装为 Agent Skill。
2026-09-27
czg86389-hub/muse2api
B
把 Muse(muse.ai) 逆向封装为 OpenAI 兼容接口，支持对话、文生图、文生视频/图生视频、多账号池轮转与 48h 自动续期。OpenAI-compatible API for Muse.ai with Chat, Image & Video generation.
2026-09-24
BerriAI/liteagents
A
A provider-independent agent SDK with the same query() interface as the Claude Agent SDK, allowing you to use the right model for every turn.
2026-09-24
mariagorskikh/talking-head-reel
A
Claude Code skill: turn a messy phone recording into an edited vertical reel (best takes, captions, animated overlays, memes)
2026-09-23
jev-chat/jev-chat-windows
B
微信（Windows 4.x）旁挂的回复辅助：窗口截图 + 本地离线 OCR 读对方消息 → Jev 判断意图 → 3 条候选一键填入，发送永远手动
2026-09-07
FeiZhuLulu/real-api-pricing
B
真实 API 定价：订阅月费 ÷ 实际可用 token，含额度、单价与分榜帕累托图。Real API pricing: subscription fee ÷ usable tokens, with allowance, unit-price, and leaderboard Pareto charts.
2026-09-07
pcm469768-droid/llm-api-speedtest
A
Benchmark LLM APIs, OpenAI-compatible gateways, and API proxies by token speed, TTFB, latency, and success rate. Compare multiple providers with reproducible JSON reports and scheduled GitHub Actions.
2026-09-05
0nicell/data-deletion-skill
A
A Claude skill for running structured personal-data erasure campaigns — 328 data brokers, OSINT tools and breach-search engines with verified privacy contacts, request templates for 13 privacy regimes, and escalation paths to 21 regulators.
2026-08-25
sunchaokun/pptx-designer
B
Code-first Python library for generating editable PowerPoint presentations with native PPTX objects, SVG support, and LLM-friendly build APIs.
2026-10-26
tenhkspark/secbrief
A
A daily vulnerability briefing for AI-agent-managed machines — deterministic matching against OSV + CISA KEV; the LLM only explains. MIT
2026-10-17
Tongyi-MAI/OpenGCUA
A
An open-source stack for generalist computer-use agents, combining code execution, GUI interaction, and cross-device control.
2026-10-14
glmn-ai/vibereceipt
A
Print a store receipt for your AI coding: tokens, what they'd cost at API prices, and how many times the agent said "You're absolutely right". Claude Code, Codex, OpenCode. 100% local.
2026-10-13
recepyagmur/EnglishPractice
A
Yerel, çevrimdışı ve taşınabilir İngilizce pratik uygulaması (Türkçe konuşanlar için): yapay zekâyla sesli/yazılı sohbet ve anında düzeltme, A1–C1 dersler ve sınavlar, Türkçe sesli öğretmen, kelime ve kalıp çalışma, YDS tarzı süreli testler. Windows ve Linux. Ollama + Gemma 4, Whisper, Kokoro, EMA Lightning.
2026-10-08
nachisama/ai-data-extractor
A
Free open-source extractor for AI coding assistant chat histories. Supports Claude Code, Cursor, Windsurf, Aider, Cline/Roo Code, and more.
2026-10-07
bottomless/agent-duel
A
Blinded battles between coding agents: two agents, one prompt, you pick the winner. Desktop app, runs locally.
2026-10-07
jinrunsen/claude-my-privacy
A
MyPrivacy: local name and company pseudonymization for Claude Code, with restoration for tools and assistant reply display.
2026-10-05
MohammadHaishemKhawaja/hashyy
A
A full, unpruned 177B MoE model at 11.5 tok/s on one 12 GB GPU. Expert streaming for llama.cpp, with the measurement harness and every dead end.
2026-10-04
whirlchat/whirl
B
The AI chat app that sweats the details. Every top model, real memory, living documents, and your own tools.
2026-09-23
NoizAI/YuE2-Turbo
A
Fast, concurrent inference for YuE2. Same model and recipe: 1.68× faster per song, 3.31× more songs per GPU.
2026-09-20
sutro-sh/jev-align
B
Build calibrated AI Functions from human feedback using Jev and GEPA.
2026-09-13
ccompactor/ccompactor
A
TypeScript CLI: extract any coding-agent session (Claude Code, Codex, Pi) into a compact, verified, provenance-linked handoff another agent can continue from.
2026-09-11
yudaprasetya007/routeVSCODE
B
⚡ Zero-reload dynamic AI model switcher & local proxy for VSCode Copilot Chat via 9Router
2026-09-09
SpinnerAppreciate/glm-flash-offline-client
B
Offline desktop client for GLM-5.3-Flash — run private, local chat sessions without a server connection.
2026-10-28
xiaoxiao150/vram-switcher
A
GPU coordination for local AI companions on AstrBot / DeepSeek Harness: persona LLM, IndexTTS voice and ComfyUI selfies take turns on one 12 GB card. No proxy.
2026-10-26
alrojo/fp4Attn
A
CroissAttn: FP4 attention forward and backward, with analysis of stochastic rounding, zero-sum softmax gradients, and sampled centering.
2026-10-12
zencefilefendi/kadi
A
⚖️ Kadı — a calibrated judge in front of your frontier LLM. Routine decisions answered locally with a conformal-prediction guarantee (≤ α disagreement with the oracle); everything else escalated and learned from. Zero-dependency Python.
2026-10-03
juliankang4/spark-scope
A
Read-only dashboard and rack panel for DGX Spark / ASUS GX10 / MSI EdgeXpert nodes running vLLM or SGLang
2026-09-25
rudratoshs/buried-injections
A
🛡️ Regex catches 0%, Meta's Prompt Guard 2 catches 1% of 629 realistic AgentDojo injection attacks when they're buried in tool output. Reproducible benchmark.
2026-09-23
deepopen-com/deepopen
B
非自回归System 1决策引擎，专为结构化类型决策场景设计 DeepOpen Multilingual, non-autoregressive System 1 decision engine.
2026-09-23
yoheinakajima/glance
A
Ask an open vision-language model typed questions about an image and get probabilities back, on your own machine.
2026-09-22
oil-oil/video-editor
A
智能通用视频粗剪与口误废案剔除 Skill：自动检测语音停顿与空白，结合大模型并发滑窗识别口误重讲与废案，硬件微淡化导出无爆音成片。
2026-09-22
qghjw1216-netizen/job-filter-assistant
A
BOSS直聘 / 前程无忧通用的岗位筛选助手 Chrome 扩展（MV3）：被动抓取岗位、AI 画像与逐岗位匹配分析、多协议 AI 接入、批量分析导出。数据仅存本机。
2026-09-15
mpociot/claude-siri-ai
A
Experimental macOS 27 App Intents model delegation provider backed by Claude Code
2026-09-13
MZULALI/toolhost
A
Let an LLM write its own tools at runtime: parsed with acorn, versioned in SQLite, run in a supervised worker process
2026-09-10
chenkb98/voxweave
A
Streaming speech language model plumbing: audio frames, turns, conversation schemas, batching, and adapters.
2026-09-04
cyptnpassira/PromptLedger
A
PromptLedger - version prompt templates, run deterministic comparison snapshots, score candidate responses, and ship diffable experiment reports. Python core plus a Go companion CLI.
2026-10-24
Addition2000/mcd-diet-guard
A
帮有忌口的朋友在麦当劳安心点餐。 不吃猪肉、吃素、海鲜过敏、乳糖不耐、不吃辣、不吃酸黄瓜……说一句话，它就查遍整家门店的实时菜单，告诉你哪些放心点、哪些“去培根”就能吃、套餐里该换成什么，算好价再帮你下单。
2026-10-16
PaddlePaddle/FlashMask
A
Efficient Mask-Aware Distributed Attention for Long-Context Training
2026-10-13
alibaba/qz-browser-agent
A
Local-first AI agent for Chrome: browser automation, web research, and data extraction with your own model.
2026-10-12
xiaomoBoy/pi-durable-assistant
A
可自托管的持久化个人 Agent：进程重启后恢复任务，支持子 Agent、定时计划、长期记忆、网页界面与微信通道
2026-10-05
obie/auto-handoff
A
Claude Code plugin that writes a handoff document before an idle session's prompt cache goes cold
2026-10-04
sebastianhaba/book-to-course-skill
A
Claude skill that turns a book (EPUB, PDF, HTML, DOCX, Markdown) into an interactive beginner-friendly course: step-by-step lessons, animations, quizzes, chapter tests, flashcards and coding exercises checked by unit tests, with progress saved to files.
2026-09-30
huangbai-AI/one-prompt-video-skills
B
一句话生成完整口播视频：写稿、口播数字人、B-roll、HyperFrames 合成、抽帧检查的 Claude Code Skills
2026-09-29
B0yko/taskdistill
A
Distil an expensive LLM API call on a narrow task into a small local model: capture traffic, curate, LoRA fine-tune with MLX on Apple Silicon, evaluate against the teacher, and serve an OpenAI-compatible cascade that escalates low-confidence requests.
2026-09-28
Vinyl-Davyl/hermes
A
Local-first handoff CLI that packs unfinished work on your disk and hands it to the next coding agent: Claude, Cursor, Codex, Antigravity, and more. No plugin. No cloud.
2026-09-28
sfmqrb/gutcheck
A
grep for meaning: ask a question about lines, logs, diffs or JSON and explore the answers live. Runs locally on CPU, one Rust binary.
2026-09-23
mpnikhil/dev-0.4b
A
Dev (dev-0.4b): a 0.4B bidirectional decision model on ModernBERT-large for routing, yes/no verification, and 1-5 rating. Non-generative single forward pass with calibrated confidence.
2026-09-19
tamaratran/jev-pruner
B
Claude Code plugin: trim long Bash output with TypeSafe Jev before the model sees it
2026-09-05
truespar/paddock
B
Native Rust inference server for open models on NVIDIA GPUs. OpenAI- and Anthropic-compatible APIs, GGUF + safetensors, FP8/NVFP4/MXFP4/Q8/Q4, built-in Studio
2026-09-04
codejunkie99/fable-orchestrator
B
Fable 5.1 orchestrates. GPT-5.6 Luna and DeepSeek V4 Flash implement.
2026-09-03
dreamers-laboratory/ai-evaluation-framework
B
Accuracy, latency p95, and cost benchmarking for model-based solutions, with per-field ground-truth scoring.
2026-08-25
AgriciDaniel/youtubepro
B
Local-first YouTube research, grounded AI insights, script writing, and thumbnail creation.
2026-10-29
oxnw/ai-phone
A
A SIP phone daemon with local speech and language models on Apple Silicon.
2026-10-24
liyupi/dafeiyu-cleaner
A
DeepSeek 蓝色大肥鱼消除计划：识别并清理 B 站（Bilibili）评论区中由 DeepSeek 大模型、AI Agent 驱动的机器人账号（大肥鱼）的 Chrome / Edge / Firefox 浏览器插件。离线规则 + DeepSeek API 精判，AI 评论检测、空间主页鉴定、鲸鱼娘一键清场。DeepSeek AI bot detector for Bilibili comments.
2026-10-23
nothinx/naskah-id
A
Skill agen AI untuk menulis dan menyunting karya ilmiah berbahasa Indonesia yang natural, baku (EYD V, KBBI), dan jujur.
2026-10-13
sepehr071/awesome-iran-mcp
A
19 open-source MCP servers (308 tools) that give AI agents the Iranian web: Digikala, Snappfood, Jabama, MrBilit, Paziresh24, Taaghche and more. Prices in Toman.
2026-10-10
Disdjj/intelligent-ui-demo
A
拆解 ChatGPT Intelligent UI 的实现，并与 AG-UI、A2UI 对照；附可运行复刻（流式 DSL 编译器 + iframe/Worker 沙箱）
2026-10-09
Odyphus/MathModel-Copilot
A
面向国赛 CUMCM、美赛 MCM/ICM 的数学建模 AI 助手与 Agent Skill：审题、模型比较、Python 运行、结果核验、论文辅助和团队协作。Mathematical modeling AI copilot for CUMCM and MCM/ICM, with reproducible workflows and a local workbench.
2026-10-05
raphaelmansuy/edgextract
A
Turn markdown into a knowledge graph with an ontology you write and a cutoff you own. Closed yes/no decisions, not free-form JSON.
2026-10-05
saawant12/ps5-ai-cli
A
Run OpenAI Codex on PS5 with controller text entry and a browser terminal for your phone or computer. Experimental homebrew.
2026-10-04
Yamz-Labs/kyojin
A
Kyojin: the Yamz inference engine for AMD Strix Halo (ROCm, gfx1151), built on ExLlamaV3. Runs 300B-class MoE models on one 128 GB mini PC.
2026-09-26
Xuanwo/cida
A
辞达：用你选择的大模型，在 Mac 的任何地方翻译和润色文字。Translate and polish text anywhere on your Mac.
2026-09-26
Show HN: Jev Plays Pokémon Red
B
Hey HN! Wanted to share a fun project I've been hacking on. Given Jev can make decisions really fast (but not fast enough to play Doom yet sadly), I wanted to try and push it to play a more complex game than Tetris. So I went with Pokémon. I've spent endless hours playing this game as a child so building this was a ton of fun. I open sourced everything in case you want to hack on it yourself here: https://github.com/christianmat/jev-pokemon The game is being streamed live including the tokens and cost - hopefully we get all the badges and don't get stuck in a cave :)
2026-09-24
amitshekhariitbhu/ai-engineering-course
B
AI Engineering Course - A free and complete AI Engineering Course to learn AI Engineering step by step - from Machine Learning, Neural Networks, and Transformers to LLMs, Fine-Tuning, RAG, AI Agents, LLM Inference, Evaluation, AI Safety, and AI System Design.
2026-09-21
fedorchenko-juli/job_search_tool
A
Automated job search pipeline: fetches roles from company ATS APIs and aggregators, filters and dedupes them, optionally scores fit with Claude, and reviews results in a local web board.
2026-09-21
ghuntley/underclass
B
underclass: an OpenAI-compatible pooling proxy that pins sessions to one account (prompt cache stays warm), cools quota-exhausted subscriptions until their window resets, and fails fast with the earliest Retry-After when the whole pool is dry.
2026-09-18
bnsd55/jevmlx
A
Jev-style parallel constrained decisions for any MLX model on Apple Silicon. Typed, schema-valid JSON in one forward pass.
2026-09-17
zorrobyte/rimagent
A
An LLM that plays RimWorld and rewrites its own playbook between colonies. RimBridge mod + self-improving agent.
2026-09-07
Datura-ai/lium-localmaxxing
A
Reproduce Lium's LocalMaxxing leaderboard runs (7 boards, 20 Verified) on a rented Lium GPU in one script — vLLM/llama.cpp recipes, measured cost per million tokens
2026-09-05
affirmitv/cc-readback
A
Read your Claude Code sessions from Claude Desktop. A local, read-only MCP server over ~/.claude with secret redaction.
2026-09-02
H0nGzA1/testpilot
A
Agentic E2E testing platform — natural-language test cases executed by a browser agent in real Chrome, judged by an LLM with video/screenshot evidence
2026-08-28
MiaAI-Lab/GLM-5.3-Flash-NVFP4-Dual-DGX-Spark
B
2x DGX Spark recipe for LibertAIDAI/GLM-5.3-Flash-NVFP4 (NVFP4, Ray TP=2, multimodal)
2026-10-27
hieu-lee/facebook-cli
B
fb: fast, token-efficient Facebook + Messenger CLI for LLMs, driving your own logged-in Chrome
2026-10-06
LAGcomcom/zen-gate
B
把 OpenCode Zen 免费模型装进你所有的 AI Agent —— Windows 托盘网关，OpenAI/Anthropic 兼容，自动适配 10+ Agent，限流自动切换
2026-10-06
THUROI0787/absent-author
B
Research and writing can be automated. Authorship cannot. Evidence list + agent skills for spotting AI-produced papers no human stood behind.
2026-09-26
amitshekhariitbhu/ai-system-design
B
AI System Design - Learn how to design AI systems built on LLMs, RAG, and AI Agents step by step.
2026-09-26
architectds/collabosm
B
Run Qwen3.8-Flash-Next (125B-A6B MoE) on ONE Colab A100-80GB High-RAM: pinned runtime, weights from HF, OpenAI-compatible endpoint, measured 2.8k-3.9k t/s prefill / 90-97 t/s decode, plus a verified pinned-RAM KV tier.
2026-09-23
Calcium-Ion/AstrLink
B
A local AI gateway for AI agents, unifying your subscriptions and API providers with smart routing and on-device privacy protection.
2026-09-19
TianyuCodings/NanoJev
B
A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.
2026-09-18
devanshbatham/commit-miner
B
Classify Git commit diffs and messages with Jev. Bug fixes, security fixes/CWEs, and change types.
2026-09-17
donvito/agent-monitor
B
Visualize your Codex and Claude Code subagents. Advanced traces view, tokens breakdown, and cost usage is also supported.
2026-09-16
wushi2333/dsh-computer-use_codex-style
B
Codex-style Computer Use for DeepSeek Harness: the official window2 13-method desktop surface plus Chromium automation, a visible overlay with Esc interrupt, per-app approvals and bundled skills. Windows, MIT, no Codex required.
2026-09-15
yifanzhang-pro/FlashREINFORCE
B
FlashREINFORCE: Critic-Free, Single-Rollout, Asynchronous RL for Agentic Language Models
2026-09-08
pfz14/chatgpt-model-meter
B
Chrome MV3 extension for chatgpt.com: shows 5h/7d quota, and checks the requested model against the model_slug the server's STE telemetry reports as actually executed.
2026-08-28
rexadbapp/io
B
A tiny, self-contained LLM coding agent. One binary, 180+ providers, zero runtime dependencies.
2026-08-27
seanyourhighness/vllm-sm12x-nvfp4-dflash2
B
All-NVFP4 vLLM v0.27.1 + DFlash2 K7 speculative decoding for Blackwell (SM120/SM121): NVFP4 target/draft/KV, 262K context, optimized language-only serving.
2026-11-09
AmSach/maple
B
MAPLE: Manifold-Activated Pruning of Latent Experts for Sub-2-BPW MoE LLMs
2026-11-02
Balanov/uncut
B
The lossless, headless coding agent harness. One 7 MB static binary, four tools, plain stdin/stdout, no silent cuts.
2026-10-26
joelhooks/lockjaw
B
🔐 Egress credential injection for agents: agents hold placeholders, the cell fills real keys at the edge for approved hosts.
2026-10-21
gojko/lights-on-software-factory
B
Lights-on software factory boilerplate for Claude Code. Based on the setup developed for Votito.
2026-10-21
openai/MultiLangBench
B
An eval based on SQLite that benchmarks language models in multiple languages.
2026-10-10
Show HN: Apogee: Rebuilding Mozilla's Orbit, fully local and private
B
Hey everyone! Last year, Mozilla released Orbit, an AI-powered browser summarizer hosted on a GCP server. After people started digging into the extension, they discovered things like backend endpoints such as store_result. Eventually, Mozilla discontinued the project. For the past month, I’ve been trying to rebuild Orbit from scratch, but with one major difference: Apogee is fully local and privacy-focused. Apogee doesn’t send or store your data. It can directly connect to your local Ollama instance for inference. I’ve also added WebGPU integration for Chrome and Transformers.js for Firefox to provide faster, local responses. It can summarize: Articles and websites, YouTube and Billie videos, Wikipedia articles ,Hacker News and Reddit threads. Install Apogee: Chrome: https://chromewebstore.google.com/detail/apogee/pgemlpomhkdc... Firefox: https://addons.mozilla.org/en-US/firefox/addon/apogeeext/ Obviously it is far from complete. Would love to hear your feedback and suggestions!
2026-09-27
charlesdove977/claude-x-jev
B
Fast, cheap, typed decisions for Claude Code. Jev (TypeSafe's decision model on OpenRouter) sorts, checks, scores, gates and verifies at 0.3s and a fraction of a cent per item. Claude keeps the reading, writing and judgment.
2026-09-24
DaVVinCi/llm-multi-agent-game-evaluation
B
A dynamic multi-agent game framework for evaluating LLM strategic reasoning, communication-driven decision-making, deception, and social-norm behavior across RPS, Hawk-Dove, and Liars Bar environments.
2026-09-24
darkdarkcocoa/codex-routing-detector
B
Find out which model actually answers your Codex requests (detects silent model substitution, e.g. gpt-6-astra served by gpt-5.6-luna)
2026-09-22
disler/self-compact-pi-agent
B
A Pi coding-agent extension that lets a long-running agent watch its own context and compact itself
2026-09-18
liyupi/ai-model-world
B
AI 大模型世界，把 556 个大模型拟人化成像素小人的可视化站点。进来就能看到此刻谁最聪明、谁最会写代码、谁最便宜、谁刚发布，往下是国内与国外分区的厂商广场、完整的发布时间线和多维排行榜。搜索认模型名、厂商和能力，输入「多模态」会直接列出全部多模态模型。数据取自 Epoch AI、models.dev、LiveBench 与 Hugging Face，每小时自动同步，所有文案由真实数据生成，不调用任何 LLM。Next.js 静态导出，零后端。
2026-09-12
Field-Logic-Ltd/ClaudeStatsBar
B
A Claude Code status line showing how big your session has got and what the next turn costs before you type
2026-09-10
bluesky585/nibble
B
A Go library for splitting long text into retrieval-sized chunks for RAG.
2026-09-08
lucidrains/declarative-attention
B
Implementation of the procedure in Language Models Can Control Their Own Attention, from Namgyu Ho et al. of KAIST AI
2026-09-05
mpaymenremora/CacheCraft
B
Semantic cache strategy bench: replay recorded request traces through candidate caching policies and measure which one actually pays off.