#open-weights

NVIDIA's Nemotron 3.5 Transcribe, released October 28, against Whisper large-v3 and Parakeet TDT 0.6B v3 on licence, languages, size, hardware and WER.

Mistral released Large 4 weights on October 29 and lists $0.70 input and $2.10 output per million tokens. Direct prices compared with DeepSeek, Kimi and GLM.

Bria FIBO 2 uses structured image controls. Compare its non-commercial weights, metered self-serve plans and separate Enterprise deployment terms.

Two US open-weight models in two days: Reflection Beam and Meta Muse Spark Open compared on licence, hardware, shared benchmarks, price and best fit.

DeepSeek's V4.1 Flash 1017 ships MIT weights and shorter answers; its new Search API bills Claude Code and Codex calls from October 31. Prices and limits.

Reflection's first model, Beam, ships 501B open weights under Apache 2.0. What it takes to run, what its benchmarks show and where you can call it today.

Black Forest Labs' FLUX 3 Image Klein: 5.6B open weights that generate and edit images on an 8GB GPU in INT8. Licence limits, API price and who it suits.

Meta's Muse Spark Open puts 816B-parameter open weights behind coding agents. What the licence allows, the hardware it needs and where you can run it.

Meta's new open-weight model against Kimi K3 and GLM 5.3 on shared benchmarks, API price, licence triggers and hardware, and which suits your work.

Fish Audio's Drama 3 takes stage directions in brackets, edits single lines in place and ships open weights. Price, licence limits and who it suits.

Compare Drex v2.0 local hosting and API access using verified limits, weight sizes, license boundaries and a practical plan for testing agent decisions.

Evaluate Step 5 Preview for agent work: hosted access, dated OpenRouter prices, context limits and the distinction between API use and planned weights.

Assess Mistral Large 4 for agent workflows: preview access, hosting choices, context limits, evaluation design and the promised open-weights release.

Cloudflare's Clef and Amazon's Decider 2B put decision models like closed Jev into open weights. Licence, size, context, latency and benchmarks in one table.

Ming-Image-0.1-Design makes UI, poster and infographic images with transparent backgrounds; its Layer variant splits a flat design into editable layers. MIT.

DeepSeek published DeepSeek-V4-Flash-Vision-Exp's weights under MIT on August 31, ten days after the API-only launch, with a reference implementation attached.

Z.ai published GLM-5.3's 753B weights on August 28 under a bespoke licence, two days after shipping GLM-5.3-Flash under plain MIT. The clause that differs.

In the 2025-26 record we checked, two labs gave a public reason before holding back open weights. What a stated delay tells a buyer, and what it does not.

fal says H3 Max renders five seconds of video in under three. The claim is the vendor's own, and the tracker it cites backs the quality half only in part.

Alibaba open-weighted Qwen3.8-Flash-Next on August 26. The hosted Qwen3.8-Flash is a documented, different artifact — licence, context and tools all diverge.

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.