#open-weights
0 tools tagged open-weights, 8 news items
News tagged #open-weights
- Cohere
Cohere's Tiny Aya L2-Thinker reasons in any language
Cohere Labs shows reasoning in the prompt's own language through data mixing: Tiny Aya L2-Thinker (3.35B) reasons that way over 93% of the time, across 60 languages.
- Hugging Face
Falcon-Emirati-7B adapts a model to a single dialect
TII's Falcon-Emirati-7B is a dialect specialized model built on Falcon-H1-Arabic for Emirati vocabulary, tone and cultural context, evaluated by native speakers.
- Mistral AI
Mistral previews Large 4, a 1T open weights model
Mistral Large 4 enters public preview with 1 trillion parameters and 49B active, trained on 3,800 Grace Blackwell GPUs in Europe. Weights ship by end of October.
- Allen Institute for AI
Ai2 open-sources AstaBrief, a fast cited report model
Ai2 released AstaBrief 8B under Apache 2.0 with training data: 51.1s per cited report against 178.5s, and its best quality win came from one cheap filter.
- Meta
Meta extends its loss of control rules to RL training
Meta's updated Superintelligence Scaling Framework now requires sandboxing and tamper-proof logging before any RL run, and adds an open-weight release section.
- Anthropic
Anthropic red teams GLM-5.3's cyber reach
Anthropic's Frontier Red Team found GLM-5.3 built working exploits where other open models scored zero, and that its safeguards fall to a deceptive framing.
- NVIDIA
NVIDIA's Kumo Tabular is a one pass tabular model
NVIDIA released Kumo Tabular, an open foundation model for tabular prediction that runs a single forward pass with no training, tuning or feature engineering.
- H Company
H Company opens Holo4, a 27B model for every interface
H Company's Holo4 27B scores 85.2% on OSWorld at $0.08 a task and works through GUIs, code, MCP and APIs at once, with every benchmark trajectory published.