Thinking Machines Lab on 5 November 2026 introduced Inkling 2 Small, a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, image inputs, variable effort and a context window up to 1 million tokens, according to the company’s news post and model card. The model is available now on Tinker with closed weights. Thinking Machines says it intends to release full weights under Apache 2.0 next week after external safety testing and red-teaming.
The model card describes a 48-layer decoder-only transformer with a sparse MoE feed-forward backbone — each token routed to 8 of 384 experts — and a vision transformer that projects images into the same hidden space as text. Output is text only; the card does not list audio.
What Thinking Machines reports for training and scores
The news post says initial pretraining (H1) covered 8.4 trillion tokens on NVIDIA GB300 NVL72 racks, running natively in NVFP4, followed by further phases including higher-quality data and long-context agent trajectories. A joint reinforcement-learning run then completed more than 17 million rollouts over nine days on 2,048 GPUs, the company says.
On benchmarks, Thinking Machines asks readers to treat scores as company-reported or third-party-labelled. The news post cites an Artificial Analysis Intelligence Index score of 31. The model-card table lists AA Index (v5.0) at 30.6 for Inkling 2 Small. Both figures appear in Thinking Machines’ own materials dated the same day. Selected coding and design numbers the company highlights at effort 0.99 include FrontierCode v2.0 at 25.4%, DeepSWE v1.1 at 72.1%, and DesignArena Website at 1340. Evals, the news post says, run at effort 0.99 and temperature 1.0 with a 1M-token maximum trajectory limit.
How Inkling 2 Small sits beside other open-weight releases
TSN has covered other large open-weight or preview releases including Ricursive’s Ripple-65B preview, Meta’s Muse Spark Open, and Anthropic’s Claude Haiku 5.5. Ripple and Muse Spark Open put weights in users’ hands on day one. Inkling 2 Small’s path is the reverse of that pattern for now: API and fine-tuning on Tinker first, with an Apache 2.0 weight drop promised after outside safety work — a schedule the company itself flags as intent, not a completed upload.
Thinking Machines also says early checkpoints ran on OpenCode under the name Fledge-Alpha to test real-world coding, and that vision plus agentic training is meant to support iterative website, game and 3D design loops with visual feedback. Those demos are company showcases, not independent audits.
What a closed-now, open-next-week MoE changes for builders
For developers, the near-term product is Tinker access: playground sampling plus fine-tuning and sampling APIs, with checkpoints downloadable when weights ship. The 594B/17B MoE shape and 1M context are the scale claims that matter for agent harnesses; the closed-weight window is the compliance and reproducibility caveat that matters until Apache 2.0 lands.
The AA Index split — 31 in the news narrative, 30.6 in the card’s v5.0 table — sits inside the same launch package, so leaderboard comparisons should cite which Thinking Machines page they used. Builders who need reproducible local runs will wait for the Apache 2.0 drop; those who need API access or fine-tuning can start on Tinker under the current closed-weight terms. That split matches Thinking Machines’ “safe path to open weights” framing: ship capability behind a hosted surface first, then open the files after outside red-teaming finishes. The company says full weights under Apache 2.0 are planned for next week after external safety testing; that remains its stated schedule until a public weight release appears.
Sources
- Introducing Inkling 2 Small (Thinking Machines, 5 November 2026)
- Inkling 2 Small Model Card (Thinking Machines, 5 November 2026)
Related stories
- Ricursive Releases Ripple-65B Preview Open Weights
- Meta Releases the Weights for Muse Spark Open, and Apple Shows It Running Across Two to Four Mac Studios
- Claude Haiku 5.5: Anthropic’s Small Model Gets Much Cheaper and Much Better at Using a Computer
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
