Kolibri-1 serves 1M tokens but was trained to 256K
Aleph Alpha released Kolibri-1 on 3 October 2026 under Apache 2.0. The headline says 1M tokens of context. The model card says it was trained on sequences up to 256K, ...
Tag: open-weights
MiMo-V2.6-Pro was quietly retrained to stop looping on tool calls
Same model names, new weights. Xiaomi retrained MiMo-V2.6-Pro and Flash to stop repeating tool calls, shipped it to the API first and the MIT weights two days later. We read ...
Qwen-Image-2.1 brings the weights back, but not the Apache licence
The Qwen image line has downloadable weights again: a 7B transformer, an 8B text encoder and an RGBA VAE. The licence is research or evaluation only, a first for the ...
Does Bonsai 2 27B really keep 98% of Qwen3.8 in 5.95 GB?
PrismML released Ternary Bonsai 2 27B on 17 September 2026, a 1.72-bit version of Qwen3.8 27B that fits in 5.95 GB. We read the model card, the whitepaper numbers and ...
Atria Dawn Preview is 744B under MIT, and the BF16 weighs 1.5 TB
No launch post and no price list, but 1.5 TB of MIT weights on the 744B GLM-5.2 base, a 16-row benchmark table and a text-only API. We read all of ...
DeepSeek V4.1-Flash arrived, and the V4 Pro retirement lasted a day
DeepSeek shipped V4.1-Flash on 10 September with a new encoder-decoder design, native vision and MIT weights, and said V4 Pro traffic would be routed into it from 14 September. The ...
Qwen3.8-Flash-Next runs 6B active and weighs 173 GiB
Somebody on the team read "6B activated" and asked which of our boxes could run it. Fair question, wrong number: the FP8 checkpoint for Qwen3.8-Flash-Next is about 173 GiB of ...
Tencent’s 770B Hy4 tops one benchmark row in 46
Most labs publish the launch chart and bury the table. Tencent did the opposite on 28 August, and the table is the interesting half. Hy4 preview is out under a ...
Nobody has confirmed the $12.9B Hugging Face report
Wednesday night, one line in The Information, and by Thursday morning every model your CI pulls at build time might be coming from a company that sells GPUs. Nvidia has ...
GLM-5.3-Flash was Ox Alpha, and the price doubles
We spent last week telling people not to send work code to Ox Alpha, because nobody would say who was running it. That question closed on 26 August: it was ...