The lab's second generation ships as a 594B mixture of experts on Tinker, with open weights waiting on…
Baidu published the raw pre-training weights of ERNIE-5.5-Flash-Base on Hugging Face, a 280B mixture-of-experts checkpoint that has never…
Joey-1 returns a calibrated probability instead of prose, costs a cent per million tokens and ships as open…
Mistral has published the full weights for its trillion-parameter Large 4 model, arguing a downloadable frontier model is…
Mirendil says Fern 0.1 was trained for under $10M with agents writing more than 99 percent of the…
REKI keeps the last three frames inside its action expert using 32 register tokens, cutting arm oscillation from…
The Chinese lab shipped open weights for V4.1-Flash-1017 and opened its in-house search engine as a paid API.
Microsoft's research group says its new MAI-Base-2 base model beats the best open rivals while using far less…
The open-weight release brings a 128K context window and doubled vocabulary to agents that run without a cloud…
Cloudflare's first in-house models, Clef and Clef-flash, return probabilities instead of generated text.
BottleCap AI has released a Qwen fine-tune that thinks 37.2 percent less for under a point of accuracy.
A 340M-parameter classifier from Fastino returns typed answers with confidence scores for agent pipelines.