JetBrains’ open Mellum2.1 uses 2.5B active parameters. New training alone lifted its SWE-bench Verified score from 2% to 47%, self-reported.
Tag: efficiency
Anthropic’s Haiku 5.5 adds a 1M context and effort controls. It’s 90% cheaper than Haiku 4.5 below 100k tokens, 50% cheaper above.
Reflection AI announced Beam, a 501B open-weight model with 23B active. Apache 2.0 is promised; the weights are due later this month.
Aleph Alpha released Kolibri, an Apache-licensed German-English model: 78B parameters, 3.46B active per token, trained to hold back when unsure.
Cloudflare open-sourced Clef and llama.cpp is adding support: models that pick from your options instead of generating text.
A solo developer fine-tuned Qwen3.8-27B so low, medium and xhigh effort settings behave in order. Self-reported gains, one seed, Apache 2.0.
A seven-month-old Beijing startup released a 309B open-weight model — built on Xiaomi’s MiMo-V2.5, with no full-attention layers anywhere.
Liquid AI’s new 280M draft model speeds up its vision model up to 3.13x — but only decoding, and only free under a $10M revenue cap.
Hugging Face’s Transformers library can now load and run llama.cpp’s GGUF quantized models directly, matching its speed on Apple Silicon.
Anthropic’s CEO called for slowing AI capability gains on 12 September. Ten days later, both labs shipped faster, cheaper frontier models.