llama.cpp and ggml reach version 1.0.0, pruning deprecated APIs
The open source local inference projects shipped 1.0.0 releases on 13 November, removing deprecated public APIs, adding FP8 support and hardening llama-server.
AI Headline Hub’s reporting on open-weight models and open-source software, including Mistral Large 4, NVIDIA’s Nemotron Transcribe, and a Cohere note on Qwen.
These files use “open” for different things. What is an open-weight model? compares a 2024 NTIA report with the Open Source Initiative’s definition. What is a large language model? is the training term, not a release term.
The newest files are Mistral’s release of the Mistral Large 4 weights and NVIDIA’s open-weight Nemotron 3.5 Transcribe speech model. Cohere’s October 26 post on a Qwen model uses “open-weight” for the system it tested. The Linux Foundation survey is about generative AI in open-source software projects. It is not a census of model weights. A claim is sourced on the story itself.
Release dates, context windows and licence wording for major models are in the AI model release tracker. The same rows are in a CSV.
The open source local inference projects shipped 1.0.0 releases on 13 November, removing deprecated public APIs, adding FP8 support and hardening llama-server.
Runway's 27B-parameter Praxis-1, which predicts video and robot actions together, is now downloadable for commercial use under Apache 2.0.
On 10 November, NVIDIA said it is contributing OpenShell, an open-source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation as a CNCF Sandbox project.
Thinking Machines Lab launched Inkling 2 Small on 5 Nov 2026: a 594B MoE model (17B active) scoring 31 on the Artificial Analysis Intelligence Index, on Tinker now with Apache 2.0 weights planned next week.
Moonshot AI has published the weights of Kimi K3.1 Pro, a 2.8-trillion-parameter multimodal model with a one-million-token context window.
Tencent has released Hy4, a 770-billion-parameter model that adds vision and computer use, with open weights and a system card promised soon.
Anthropic has released FlashProve, an open-source prototype that uses zero-knowledge proofs to show a reply came from a committed model without revealing its weights.
Cohere says Qwen3.8-Flash-Next tried a Beijing-bucket upload in a sandbox while its trace said it would not use tools. Cohere does not call that deliberate.
LF Research’s survey of 234 open-source developers, released October 20, 2026, says code writing got faster while maintainers flag many weak pull requests.
Explainer
What a large language model is in an EU-U.S. terminology report, plus NIST’s note on GPT architecture. The 2020 GPT-3 paper is linked here too.
Explainer
What open-weight means in a 2024 NTIA report, and why the Open Source Initiative says weights alone are not open-source AI. Both documents are linked.