Inkling-Small: 276B open-weight multimodal MoE that fits one GPU
Thinking Machines' Apache-2.0 model hits 80.2% on SWE-bench, self-hosts on one B300, and lives on Fireworks and OpenRouter. Plus LLM 0.32 and MiniMax H3 on Mac.
Copy markdown
Frontier-grade weights, single-GPU footprint
Thinking Machines Lab's Inkling-Small is a 276B-total / 12B-active multimodal MoE (text, image, audio) with a 1M-token context, out under Apache 2.0. It scores 80.2% on SWE-bench Verified and 95.5% on AIME 2026, and the NVFP4 checkpoint (180 GB) runs on a single B300 or two H200s — with day-0 support in vLLM, SGLang, Transformers 5.14 and llama.cpp/Unsloth.
No B300? Rent it by the token
Inkling-Small is already live on Fireworks AI and OpenRouter behind OpenAI-compatible endpoints — roughly $0.45/$1.20 per million tokens in/out on OpenRouter — so you can wire the open model into a coding agent or doc-understanding pipeline without owning a supernode.
LLM 0.32 adds reasoning traces and the Responses API
Simon Willison's LLM CLI shipped 0.32: it now captures model reasoning traces, speaks OpenAI's Responses API, runs server-side provider tools, and rewrites its SQLite logging — with matching plugin bumps for Anthropic, Gemini and OpenRouter. Update to log and script agent runs across providers from the terminal.
MiniMax H3 now runs on Apple Silicon
Days after MiniMax open-weighted its omni-modal H3, a community MLX port brings it to Macs: text, image and audio in, up to 15 seconds of video with sound out, all on-device. A weekend unlock for local multimodal generation without a cloud bill.