#open-weights

DeepSeek's changelog now carries a GA entry for V4-Pro across app, web, and API, MIT-licensed 0813 weights on Hugging Face, and pricing changing August 16.

GLM-5.3 reuses GLM-5.2's base model, so Z.ai says every gain is post-training. Terminal-Bench 3.0 jumped 4.6 to 28.3, and weights are promised later.

Mistral's Shieldstral reads your safety policy as prompt text at inference time. A 3B Apache-2.0 guard model with a mixed, not one-sided, benchmark record.

Alibaba promised Qwen3.8-Max and a 27B sibling for the week of 10 August. Updated: Max weights landed 8 August under a bespoke licence. Run these checks.

At the time of writing there is no Wan 3.0: no Alibaba announcement, no weights, no listing. July really shipped one open-weight model and two research papers.

Thinking Machines' 276B/12B Inkling-Small leads its 975B parent on the vendor's own coding and reasoning table, then loses badly on factuality.

OpenAI, Anthropic and Google skipped NVIDIA's Open Secure AI Alliance a month after joining Akrites. That selectivity, not the absence itself, is the story.

MiniMax launched H3, a 2K video model with native synced audio at $0.13 per second. Weights are promised within days, not published yet.

Kimi K3's weights are open, but the licence is bespoke: a $20M revenue trigger for resellers, UI attribution rules, and a 1.56TB download to plan for.

Kimi K3 is 1.56TB of weights and needs 64+ accelerators to serve. Here is the real cost maths of self-hosting a trillion-parameter model in 2026.

Tencent's Hy3 shipped Apache 2.0 at a sub-300GB FP8 footprint. A routing guide: send agentic tool-use to Hy3 and repo-scale coding to a frontier model.

Gemma 4 12B processes text, image, audio, and video with no separate encoders, fitting in ~7GB at 4-bit. A guide to running private multimodal agents locally.

NVIDIA shipped Nemotron 3 Ultra, a 550B open MoE reasoning model with weights, data and recipes under a permissive license. It runs fast but trails Kimi K2.6.