Home

Open Source

Curated open-source AI projects with verified GitHub data and hands-on breakdowns.

Sequence In, Structure Out: Why Did a 62-Star Repo Make Nature?

structRFM is a structure-guided RNA foundation model from the MIRACLE lab at USTC; the paper appeared in Nature on November 5, 2026 (DOI 10.1038/s41586-026-11210-2). The repository heqin-zhu/structRFM is MIT licensed and showed 62 stars in a November 6, 2026 GitHub API check; it was created in February 2025 and its README still points to the August 2025 bioRxiv preprint. The team says the roughly 21 million sequence-structure pairs, model weights and source code are public: the dataset is built from RNAcentral and BPfold with a 512-token length cap and is 4.5 GB on Zenodo. SgMLM structure-guided masked language modeling and MUSES multi-source structure ensembling let the model emit sequence-level, nucleotide-level and pair-level features from sequence alone, supporting secondary and tertiary structure prediction, splice site prediction, IRES identification, ncRNA classification and zero-shot homology classification.

17 Seconds on a CPU: Can an API Prove Which Model Answered?

anthropics/FlashProve is an open-source research prototype for zero-knowledge verifiable inference from Anthropic (Apache-2.0, 843 stars per a November 5, 2026 GitHub API check, version 0.16.0, officially labeled a research artifact: not maintained and not accepting contributions). The three-step protocol: the provider publishes a weight commitment ID offline (about 39 minutes on one H100), attaches a GPU-generated proof to each reply (about 19MB per 1024 tokens), and anyone verifies with cp_verify on a plain CPU in seconds. It runs Llama 4 Maverick at full 400B size on a single H100 at about 110 tokens per second of proof prefill with about 17-second CPU verification, at a cost of roughly 70 to 95 times normal inference GPU time; soundness is formally proved in Lean 4. It proves provenance, not correctness, and no major provider ships proofs yet.

Even Failed Experiments Trained It: An MIT Science LLM

On November 2, Deep Principle released and open-sourced Hydrogen-1.0: a 753B science LLM post-trained from GLM-5.3, under MIT, live on ModelScope and HuggingFace (deep-principle/Hydrogen-1.0, about 1.51TB BF16, plus NVFP4/MXFP4 variants, tested on 8 GPUs with SGLang). Vendor-reported FSR 38.3% beats GPT-6-Astra at 33.3% and GPQA-Diamonds 93.0% nears Kimi-K3; its signature is training on its own lab records including failed experiments, with a Lab Compile step that catches faulty experiment instructions.

How a Pure-C Engine Runs a 744B MoE on an Ordinary PC

JustVugg/colibri is a hierarchical MoE inference engine written in pure C (Apache-2.0, 47,567 stars per a November 4, 2026 GitHub API check): hot experts live on the GPU via CUDA or Vulkan, warm ones in RAM, cold ones stream from NVMe. Thirteen engines read published GGUF files directly, with a floor of 8 GB RAM and 22 GB disk, installed in three steps via START-HERE.bat on Windows or start-here.sh on Linux and macOS. The README demo shows GLM-5.2 at 744B MoE int4 ready in 32 seconds with 9.9 GB resident. Its engine list does not include Hy4.

Stored 32 Layers, Run 56: SenseNova Fits 27B Into 9B

On October 30, 2026, SenseTime open-sourced SenseNova Loop-9B-Preview (github.com/OpenSenseNova/SenseNova-Loop, Apache-2.0): a Looped Transformer that stores only 32 distinct decoder layers, runs the middle 24 shared-weight layers twice, and executes 56 decoder-layer computations per forward pass - about 9B parameters with roughly 14.6B of compute, which the vendor reports matching 27B-class models. A two-stream mHC keeps information flowing across passes, with 256K context, multimodal input, and a torchtitan fine-tuning recipe.

DeepSeek Open-Sources Agent Orchestration: Let AI Manage AI

On October 31, 2026, DeepSeek Harness unveiled the architecture behind its Super-agent project and open-sourced the prototype (github.com/deepseek-ai/dsh-superagent-prototype, MIT, about 9,374 stars as of November 2). The core idea ports classic systems thinking - scheduling, caching, compute-storage separation, watchdogs - into an agent: a two-tier Router plus Worker, event-driven, async multi-channel, external Watchdog. In synthetic-load experiments half of tasks finished in 12.4 minutes and 95 percent in 15.7, but at the cost of more tokens - trading tokens for time and stability.

Breeze TTS 2: Open Weights No.1, Yet Not for Sale

BreezeBlue (Shanghai, founded June 2025, 15 people; founder Yang Bin is ex-MiniMax tech partner; 6 million USD seed announced August 25) open-sourced Breeze TTS 2: repo breezeblue-ai/breeze-tts (1,230 stars, Apache-2.0 code), about 3.48B parameters, Chinese-English bilingual, weights on Hugging Face and ModelScope. On Artificial Analysis's Provider Voice arena (late-August snapshot) it scored Elo 1215, first among open weights, 90 points over Fish Audio S2 Pro (1125); but on the controlled-voices board (8 fixed voices) it ranks only third among open weights (1002, tied with Fish S2 Pro, behind Mistral Voxtral TTS at 1010) and sixth overall — "first" needs its board name. README figures: under-40ms first audio and RTF 0.32 on a warmed H100 fast path, about 7.7GiB VRAM eager (12GB card advised) or 14.4GiB with --fast-all (24GB; the fast path costs more VRAM); four-stage architecture — T5Gemma2 text encoder, Qwen3-style 28-layer backbone, 12-layer depth decoder, Mimi-style VQ (16 codebooks, 12.5Hz). The hook is the dual license: code Apache-2.0, but weights, derivative models, and self-hosted outputs are research/non-commercial — commercial use needs written approval from RESONIA or the hosted platform (34 USD per million characters); community INT8 quantization runs it on an 8GB card (one developer measured a 7.4GB peak), plus a ComfyUI node pack and a community MLX 4bit port.

Unlimited OCR 2 Open-Sourced by Baidu: 3600 Pages in One Go

Baidu released and open-sourced Unlimited OCR 2 on October 27 (repo github.com/baidu/Unlimited-OCR, MIT license, weights also on Hugging Face and ModelScope): it targets any-page-count document parsing, officially demoed ingesting all seven Harry Potter books — 3,600-plus pages — in one pass, scoring 84.6 on MPDocBench and 77.1 on the 100-page-class MarathonDocBench for end-to-end SOTA on both (technical report exact values 84.59/77.06; runner-up Kimi-K3 at just 33.74). SR-SWA (Sparse Reference Sliding Window Attention) mimics how humans copy books — looking back only when needed — with single-step decoding at 7x Full Attention and 5.6x the previous R-SWA; per the technical report, peak KV cache on 100-page tests drops from 47GB to 1.77GB (minus 96.2 percent) and 16-way concurrent throughput rises 3.65x, all self-reported pending reproduction. Three product capabilities: native layout-element segmentation, cross-page table merging, and seamless cross-page paragraph continuation. Generation one once topped four boards in a day, passed 10 million cumulative HF downloads, and earned a LeCun repost; generation two climbed to 29,757 stars in three days (measured Oct 30; official announcement said 28,500). Honesty corner: HeadingTEDS is only 48.20, so beware if you depend on heading trees; the 3B/500M-active parameter count is a third-party report the vendor has not confirmed; deployment of gen two recommends SGLang (vLLM support belongs to gen one). Division of labor with the site's six-way OCR review and GLM-OCR/DeepSeek-OCR pieces: those cover single pages to dozens; this one covers any page count.

Whisper's rival is here: NVIDIA tops the open ASR board

NVIDIA open-weighted Nemotron 3.5 Transcribe on October 28 (repo name 30B-A4B; per the model card 34B total parameters including the audio encoder with only about 4.3B active per step, on a Lightning 30B-A3B language backbone): transcription in 40 languages, translation across 31, speaker attribution in 13, one-hour audio with up to 8 speakers, and a batch-only offline design with no streaming (a separate 0.6B sibling handles streaming). Two official accuracy figures coexist: 4.71 percent mean English WER on the model card's nine public sets (versus Qwen3-ASR-1.7B at 4.94) and 5.33 on the leaderboard site (private held-out sets added since October 9); Artificial Analysis independently measured 2.81 percent, close to Voxtral Small's 2.77 despite activating roughly a sixth of the parameters. The OpenMDW-1.1 license (a Linux Foundation permissive license) allows free commercial use, redistribution and fine-tuning; weights ship in BF16/FP8/NVFP4/GGUF, with GGUF targeting NeMo-Speech.cpp though the launch blog still called quantized versions coming soon — weights posted does not mean local tooling ready. Training data Granary v2 lists 463,062 public-manifest hours across 48 languages (about 840k fully processed is a separate figure). Division of labor with the site's Hy-SimulMT interpretation piece: transcription is dictation, not real-time translation.