Open source
Models, runtimes and libraries that anyone can run and change.
15 stories, newest first
- Thinking Machines' Inkling 2 Mini loops its layers under Apache 2.022 November 2026. The open model loops 12 of its 24 layers three times per token, and a GitHub project squeezes it to 5.38 GB for Apple silicon.
- Bytedance team says three training phases are enough for AI agents22 November 2026. A paper with several US universities says stacked training pipelines hide which step helps; The Decoder calls its open model competitive.
- Ollama adds Thinking Machines' Inkling 2 Mini for Apple silicon Macs21 November 2026. Ollama says the 26B model (5.4B active) fits in about 20GB and scores 79.2% on SWE-Bench Verified. PC support is promised, not shipped.
- Google donates Agent Substrate and gVisor to the CNCF21 November 2026. The agent-lifecycle and sandboxing projects move to vendor-neutral governance; Google says agents resume in under 500 milliseconds.
- RadixArk releases DaisyRose-27B, a Qwen3.8-27B tuned with RL19 November 2026. Checkpoints come in BF16, FP8 and NVFP4; RadixArk says Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%.
- Hugging Face shows 180B MoE model running in 32GB of RAM19 November 2026. The post pairs Transformers with expert streaming and a demo Space, but the text available gives no speed figures or setup details.
- Google open-sources ABFS, a filesystem and build cache for Android18 November 2026. Google says the tool cuts Android checkouts from about 36 minutes to under 8 seconds and full builds from about 80 minutes to just over 2 minutes.
- Hugging Face moves Inference Providers onto OpenRouter's router17 November 2026. Language model requests now route across providers with retries, and custom provider keys no longer work for them, Hugging Face says.
- Meituan publishes weights for its 1.6-trillion-parameter LongCat-2.516 November 2026. The MIT-licensed agent model trails Claude Opus 5 on Meituan's harder terminal tests, and its self-hosting guide covers only Huawei Ascend chips.
- Runway releases Praxis-1 robot model with open weights15 November 2026. Code and weights come under Apache 2.0, and Runway's own tests put the model first among World Action Models on Nvidia's RoboLab-120.
- Unsloth updates Qwen3.8-27B GGUFs with Dynamic V3.1 quantization15 November 2026. Unsloth says the preview algorithm adds 5-10% accuracy at the same size, in most 2-bit to 6-bit builds but not the 3-bit ones.
- llama.cpp and ggml hit v1.0.0, with API breaks tied to major versions14 November 2026. The ggml-org team says breaking C API changes will now bump the major version, giving downstream apps a clearer signal.
- Google open-sources AQuA, a tool to trace production agent failures14 November 2026. The reference implementation clusters recurring failures and links them to deployed code. Google's 32-session replay showed fewer failures after prompt changes.
- Granite 4.1 3B and 8B models get LLMON, an LLM-native markup language13 November 2026. A Hugging Face post says the markup is now in both models, but the captured page gives no syntax, license or usage details.
- Google Cloud open-sources TPU Sync for KV cache and weight transfer12 November 2026. Apache 2.0 library moves KV cache and weights across TPU memory and network. Google reports large TTFT cuts; builds are development-only.