#open weights
15 published articles
Drama 3 puts directed TTS into open weights under a revenue gate
Fish Audio shipped a speech model you can brief in plain language and scope to a single word. The weights are public. The licence and the unfinished editing tools still decide how far that goes.
EmbeddingGemma 2 moves Google's small embedder to Apache 2.0 and adds images and audio, but text scores barely move
Code retrieval jumps 9.92 points. Google's own audio chart plots a sub-1B rival higher, and that rival bars commercial use.
Mistral Large 4 is billed as open weights, but its own AI Act papers list proprietary licenses
Mistral's 6 October preview of the 1T-parameter model promises weights by the end of October. Hugging Face had no repository at 23:18 SGT, and Artificial Analysis scores MiMo-V2.6-Pro, GLM-5.3 and Kimi K3 higher.
GLM 5.3 on Amazon Bedrock: Z.ai's 753B model reaches eligible enterprise customers through two cross-Region profiles
AWS's 5 October post adds prompt caching and OpenAI-compatible APIs, and walks through authorized security testing with Strix. It names no price and lists only US and Global cross-Region profiles.
Reflection previews Beam, a 501B open-weight MoE model with 23B active, and says Apache 2.0 weights come this month
The 5 October post opens a waitlist for early access. Reflection's own table puts Beam behind GLM 5.3 and Kimi K3 on coding rows while it pitches inference efficiency.
Perplexity Decisions API and pplx-decider-v1-27b: $0.04 per million input tokens, Apache 2.0 weights
The hosted API answers noul, choice and score questions against text, JSON or images. Docs list a 262,144 token input limit and 10 requests per second per organization.
Cloudflare Clef and Clef-flash: Apache 2.0 multimodal decision models on Workers AI, with a 64k context window
The 27B and 9B models return probabilities for typed questions in one pass. Cloudflare prices Clef at $0.24 per million input tokens and says median latency is 209 ms.
Aleph Alpha Kolibri-1 opens 78B German and English MoE weights under Apache 2.0, with 3.46B active parameters
The 3 October release needs one H200 or two A100 80 GB cards plus a vLLM plugin. Aleph Alpha's blog and model card give different German shares of the training mix.
Anthropic says Z.ai GLM-5.3 brings Mythos-class cyber skills with weak open-weight safeguards
NIST CAISI already called GLM-5.3 the most cyber-capable open-weight model to date, about four months behind the U.S. frontier. Anthropic adds 64% to 100% simulated safeguard bypass rates.
NaiveAI releases Naive-N0.5-Flash on 27 September, a 309B MIT-licensed model it says AI helped design
NaiveAI says the 309B mixture-of-experts model has 15.5B active parameters and a native 1M-token context with no full-attention layers. It lists API prices of $0.10 input and $0.40 output per million tokens but gives no launch date.
Xiaomi releases MiMo-V2.6-Pro and Flash as MIT-licensed open weights with more than 7,000 RL environments
The 1.02T-parameter Pro model scores 46 on the Artificial Analysis index, by Xiaomi's count. Xiaomi also published its RL framework and training costs. Every benchmark in the release is vendor-reported.
DeepSeek's 17 September V4.1-Flash report cuts global KV cache to 890 bytes per token and leaves the quality cost unmeasured
The 51-page arXiv paper pairs FP4 KV caching, cross-layer cache reuse and an encoder-decoder split. It reports no serving speed and no ablation for its two riskiest shortcuts.
FLUX 3 Action: open 7B world action model with LeRobot DROID and SO-101 checkpoints
Black Forest Labs released a fine-tunable world action model that maps camera frames and text to action chunks. It reports 42.92% on RoboLab-120. The weights use the FLUX Kommunity License.
NVIDIA is buying the open-model town square. The stays-open pledge is the product.
NVIDIA announced a $12,930,300,000 acquisition of Hugging Face. The 8-K adds about $11.9 billion payable to stockholders, up to $1 billion in retention equity, close expected in the first half of 2027, and required regulatory approvals. Hub neutrality is NVIDIA's public promise, not a close condition printed in the filing.
DeepSeek-V4-Pro ships on Hugging Face under MIT, not a community carve-out
arXiv 2606.19348 and the model card both cite MIT on weights and repository code. Export and deployment diligence still sit with the buyer.