Kwaipilot open-weights KAT-Coder V2.5 Dev: 35B MoE, 69% SWE-bench
Apache-2.0 weights, 256K context, run local on vLLM; Air/Pro APIs plug into Claude Code, plus open drops from Upstage Solar Open2 and Microsoft Mage-Flow.
Copy markdown
Run a 69%-SWE-bench coder on your own box
KAT-Coder-V2.5-Dev is a 35B-parameter MoE with only 3B active, Apache-2.0, and a 262K context window — it scores 69.4% on SWE-bench Verified, so you can serve a near-frontier agentic coder locally via vLLM, SGLang, or HF Transformers with no API bill and no usage cap.
Rather not self-host? Air runs at $0.15/M
Kwaipilot also sells hosted tiers: Air V2.5 is $0.15 in / $0.60 out per million tokens with prompt caching and drops straight into Claude Code and OpenHands, while Pro V2.5 ($0.74/$2.96) posts SWE-bench Pro 65.2 — second only to Claude Opus 4.8.
Upstage open-weights a 1M-context agent model
Solar Open2 250B (15B active MoE, 1M context, commercially usable license) hit Hugging Face on July 22 scoring LiveCodeBench 92.4 and MMLU-Pro 86.2 — a rare open model tuned for tool-calling across English, Korean, and Japanese, servable on four H200s (two with quantization).
A 4B image model that beats FLUX.2 on GenEval
Microsoft's Mage-Flow does native-resolution text-to-image and instruction edits at just 4B params; the Turbo build hits GenEval 0.88 — past 9B FLUX.2-Klein — and renders a 1024-square image in 0.59s on one A100. Weights are on Hugging Face and GitHub for weekend image tools.