Models & Research
bluemat briefs and decodes on Models & Research.
- Prime Intellect audit finds broken graders and reward hacks in three public benchmarks
- Qwen 3.8 27B on a laptop nearly matches GPT-6 Astra at reading model weights
- Why evicting KV cache tokens can free almost no GPU memory on vLLM
- NVIDIA releases Nemotron 3.5 Transcribe, an open promptable speech model
- Liquid AI releases LFM3, six open-weight 2B-class models built for local agents
- Study: fine-tuned rerankers beat decision models on Qwen backbones up to 9B
- OpenAI's GPT-6.1 Terra halves API input prices and targets computer-use work
- Moonshot ships Kimi K3.1 Pro via API, promises open weights by Nov 4
- Apple's 20B on-device Mac model: about 2 GiB of RAM, weak at math, scriptable
- Microsoft-Decision-1 scores fixed options with probabilities instead of text
- Recursive releases Flint, a 40B open-weight MoE coding model built to run locally
- Researchers evolve specialized agent harnesses instead of one global harness
- Mistral ships Voxtral 3: speech models plus a LiveKit-based voice agent SDK
- Cloudflare previews Clairic, a typed extraction model on Workers AI
- Inception releases Mercury 3, a diffusion LLM claiming 1,000+ tokens per second
- Muse Spark Open decoded: Meta's 816B open-weight coding model, hosted or on Macs
- Liquid AI releases d1-ultra, a decision model that returns probabilities
- Grok Voice 3.0 lands in the API at the same $0.08/min price
- Vals AI maps the case for model routers, with its own router results still to come
- Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B model for local agents
- HarnessSQL trains small Qwen3 models inside their SQL agent harness
- OpenAI's Decisions API enters public beta on gpt-6-luna at $0.10 per 1M input tokens
- Anthropic releases Claude Haiku 5.5 at a tenth of Haiku 4.5's per-token price
- Cactus Compute releases Whistle, a 16.9 MB on-device speech-to-text model
- Study: cutting coding-agent tokens with compaction can make runs 20–80% slower
- Microsoft launches MAI-Transcribe-2-Streaming, claims No. 1 accuracy on Artificial Analysis
- Cloudflare's Clef decision models: where they beat an LLM call, and where they don't
- Cloudflare's Clef: open-weight decision models for typed classification steps
- Meta and UW paper lets agents edit their own context as a file
- AgentWorld benchmark: under a third of multi-agent team actions contribute to success
- OpenAI releases GPT-6.1 Sol, claiming near-Astra results at a fifth of the token price
- Liquid AI releases d1 decision model and open-source Pipette on-device benchmark
- Cohere releases Embed 5 Pro and Fast with a shared embedding space
- Sebastian Raschka puts Jev in the context of text classification history
- Anthropic ships Claude Sonnet 5.5 at Sonnet 5's price, claiming faster, cheaper runs
- Sonnet 5.5 ships, and Simon Willison says it now powers Claude's free tier
bluemat home