Ranked by importance and recency. Items older than about 30 days drop off.Updated 28 Oct 2026, 22:00 SGT
Cloudflare's multi-stage security-audit skill for coding agents: one agent finds vulnerabilities and an independent verifier re-checks each. The repo was created in June and blew up this week. As of Oct 28: +16,388 stars this week, 51,025 total, #1 on GitHub weekly trending.
Dessa is a full-duplex speech-to-speech model trained from scratch only on human conversation audio (no text, no pretrained model). It interrupts, backchannels, and laughs. Block says it was trained on one NVIDIA GB200 NVL72 and runs locally on a high-end MacBook (M4 Max and up). It is not released: download is “coming soon”, early access is only for a small group of AI safety researchers, and no open-source release or release date has been announced. One X post had 13.49M views and 69K likes.
1T parameters, 49B active. Mistral said the weights would come by the end of the month. A community post claims Oct 31. As of Oct 28 evening it was not on Hugging Face. Treat “released today” posts as unconfirmed.
Why it matters This is an upcoming open-weights release, not a model you can download yet.
Nothing in this range.
1T parameters, 49B active. Mistral said the weights would come by the end of the month. A community post claims Oct 31. As of Oct 28 evening it was not on Hugging Face. Treat “released today” posts as unconfirmed.
Why it matters This is an upcoming open-weights release, not a model you can download yet.
Prizes are $25K, with an open-weights track. A related update is the ComfyUI OPENjutsu post.
Why it matters A wave of open ComfyUI apps is likely around the deadline.
Cloudflare's multi-stage security-audit skill for coding agents: one agent finds vulnerabilities and an independent verifier re-checks each. The repo was created in June and blew up this week. As of Oct 28: +16,388 stars this week, 51,025 total, #1 on GitHub weekly trending.
A local LLM gateway on 127.0.0.1:3425. It speaks the OpenAI, Anthropic, and Gemini APIs and converts between them, so Claude Code can use Kimi and Codex can use DeepSeek. It switches accounts when a quota runs out. MIT license. 24,419 stars as of Oct 28. No separate release date was recorded.
HKU's open-source personalized AI tutor. As of Oct 28: +2,376 stars that day, 47,336 total.
As of Oct 28: hallmark +2,109 stars that day, 40,092 total; impeccable +6,713 stars this week, 91,473 total.
A 3.3B mixture-of-experts model for long-document OCR, MIT license. As of Oct 28: Hugging Face trending #12, 300 likes, 8,357 downloads.
Dessa is a full-duplex speech-to-speech model trained from scratch only on human conversation audio (no text, no pretrained model). It interrupts, backchannels, and laughs. Block says it was trained on one NVIDIA GB200 NVL72 and runs locally on a high-end MacBook (M4 Max and up). It is not released: download is “coming soon”, early access is only for a small group of AI safety researchers, and no open-source release or release date has been announced. One X post had 13.49M views and 69K likes.
33B parameters, Apache-2.0, 512K context, aimed at agentic coding. As of Oct 28: Hugging Face trending #3, 699 likes, 7,412 downloads. The vendor reports Terminal-Bench 2.1 at 80.9 (unverified).
A new flagship for coding and long agentic work. One 1M-token window covers input and output (a single response can exceed 64K, up to about 1M). Google says it is #1 on the Vals Index. Intro price is $2 / $10 per 1M input/output tokens through Nov 30, then $4 / $20. A linked report frames it as Google's comeback after scrapping Gemini 3.5 Pro and a DeepMind leadership shake-up.
TypeSafe's Jev is a closed “decision model” that outputs calibrated probabilities instead of text, as an alternative to using an LLM as a judge. Open reimplementations exist, including Kev. The reported window is September–October, with no single day.
123456.ai is a bilingual AI intel & resource hub: daily curated news and projects, a categorized tools directory, model leaderboard snapshots — and soon, a unified model API, sandboxes and image generation.
Open-source projects and models from the current briefing, linked to their repos or homepages.
Cloudflare's multi-stage security-audit skill for coding agents: one agent finds vulnerabilities, an independent verifier re-checks each. As of Oct 28: +16,388 stars this week, 51,025 total, #1 GitHub weekly trending.
MIT CLI and MCP server that lets coding agents reverse-engineer binaries, APKs, and Electron apps. v6.8.0 on Oct 25. About 400K stars. #1 GitHub AI project by stars gained Oct 19–25.
Local LLM gateway (127.0.0.1:3425) that speaks the OpenAI, Anthropic, and Gemini APIs and converts between them. Switches accounts when a quota runs out. MIT. 24,419 stars as of Oct 28.
HKU's open-source personalized AI tutor. As of Oct 28: +2,376 stars that day, 47,336 total.
33B, Apache-2.0, 512K context, for agentic coding. As of Oct 28: HF trending #3, 699 likes, 7,412 downloads. Vendor-reported Terminal-Bench 2.1 is 80.9 (unverified).
Baidu's 3.3B mixture-of-experts long-document OCR model, MIT. As of Oct 28: HF trending #12, 300 likes, 8,357 downloads.
Three services are in the works. Leave your email and we'll ping you at launch. (Demo: the form doesn't submit anywhere yet.)
An OpenRouter-style unified model API: SDK-compatible, auto-routing with fallbacks, one bill.
Isolated environments for agents, code execution and experiments. Rent by the hour, destroy when done.
Multiple image models in one place, Chinese-prompt friendly, with style presets and batch runs.