FreshSynapse

AI trends & daily updates, in plain English

AI Models

Thinking Machines Publishes Inkling 2 Small's Open Weights, a Week After Launch

By FreshSynapse Editorial · · 3 min read

A white robotic hand reaching up against a pale blue sky

Photo: Hippopx (public domain). Illustrative image.

What happened. Thinking Machines Lab published the weights for Inkling 2 Small on Hugging Face on 12 November 2026, ending the model’s first week as a hosted-only release. The lab, led by former OpenAI CTO Mira Murati, announced the upload on X at 20:50 UTC; the files landed at 21:38 UTC. The weights arrive under Apache 2.0, in NVFP4 and MXFP8 quantisations, a week after the 5 November launch through the lab’s Tinker platform — right on the schedule the lab announced at launch.

What the model is. Inkling 2 Small is a multimodal mixture-of-experts model: 594 billion total parameters with 17 billion active per token, taking text and images in and producing text out, with up to 1 million tokens of context. Each token is routed to 8 of 384 experts, and the model exposes adjustable thinking depth — a “variable effort” control that trades speed for thoroughness. It scores 31 on the Artificial Analysis Intelligence Index (30.6 on the model card), level with Gemini 3.8 Flash per the lab. That score is vendor-reported. The model was trained on NVIDIA GB300 NVL72 systems.

The catch: it is not a model you run at home. The main NVFP4 checkpoint needs a GPU cluster with at least 500 GB of aggregated VRAM — the model card gives 2 NVIDIA B300 or 4 NVIDIA H200 GPUs as examples — and the MXFP8 checkpoint needs about 800 GB. The download itself is about 394 GB across 25 safetensors files. This is an open-weights release for organisations with serious hardware, not a laptop download. For most users, the practical route is the widened API access announced the same day: Together, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare, with serving and training integrations including Unsloth, SGLang, vLLM and TokenSpeed, plus post-training through Miles, SkyRL and Snowflake Cortex Training. On Tinker, the lab’s own platform, it costs $0.30 prefill and $1.00 sampled per million tokens.

The open question. The week’s delay was said to be for outside safety testing, but the lab has not published those safety-test results as of 13 November. That is worth noting, because the whole point of a delay for testing is the evidence it produces. The lab kept its promise on timing — the weights arrived when it said they would — but the safety case remains unseen. The release is still a significant one: a genuinely open Apache 2.0 licence, unlike Edge0-145B’s community licence, which restricts commercial serving, and a model that anyone with the hardware can download, fine-tune and build into their own products. It lands in a week when llama.cpp hit 1.0 with a stable API for local apps — the open-weights ecosystem is compounding fast. For the business behind the lab, read our report on Thinking Machines’ funding round.

Sources

  • Thinking Machines
  • Inkling 2 Small
  • open weights
  • Apache 2.0
  • Mira Murati

Related articles