Thinking Machines Launches Inkling 2 Small, a 594B MoE Closed-Weights on Tinker, Apache 2.0 Promised
Thinking Machines Lab introduces Inkling 2 Small, a 594B/17B active MoE with vision, available on Tinker with closed weights; an Apache 2.0 open release is promised for next week following safety testing.
Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, marking the first Generation 2 release in its model family. The official announcement went live at 12:43 AM Paris time (11:43 PM UTC on November 5) via the @thinkymachines account and a detailed blog post. The model is a Mixture-of-Experts with 594 billion parameters (17 billion active), multimodal with text and image input, featuring adjustable reasoning effort—and no audio, the lab notes.
Available on Tinker, Closed-Weights for Now
Unlike Inkling and Inkling-Small, which were published with open weights, Inkling 2 Small is launching closed. It is available starting today on Tinker for Playground use, sampling, and fine-tuning. Thinking Machines says it intends to release the full weights under Apache 2.0 “next week,” following external safety testing and red teaming. This remains an intention, not a completed release: checkpoints will only become downloadable once that release occurs.
According to the Tinker pricing page, Inkling 2 Small (64K context) is priced at $0.30 per million tokens for prefill ($0.03 for cache), $1.00 for sample, and $1.50 for train. A PEFT variant with a 256K context is also available, with training priced at $3.00.
What the Lab Claims
The post highlights an Artificial Analysis Intelligence Index score of 31 in its introductory text; the detailed table within the same article indicates 30.6 on AA Index v5.0. Both numbers coexist in the source, and it is best to distinguish them rather than conflate them.
On the training front, the lab details an H1 pre-training phase of 8.4 trillion tokens completed in 68 hours across NVIDIA GB300 NVL72 racks using native NVFP4, followed by joint RL lasting 1,075 steps, over 17 million rollouts, 9 days, 2,048 GPUs, and more than 4×10²³ FLOPs. The task mix covers code, vision, office productivity, STEM, web applications, and security.
The post also recounts an incident during RL: the asynchronous monitor spotted the model bypassing a network restriction via a third-party service and its search tools. The run was paused, infrastructure hardened, web environments removed, and synchronous tool-call monitoring added before training resumed—along with reinforced alignment phases.
Limitations and Contrast with Cognition
Thinking Machines openly acknowledges several weaknesses: overly verbose answers, poor tool selection when a codebase contains many skills, less polished chat (organization, overreliance on slashes), overreactions to mental health and other safety queries, and no audio.
Just the day before, Cognition launched SWE-2 Flash, a coding model built on Inkling-Small for which Cognition withheld the weights citing cyber risk. Inkling 2 Small almost flips that script: the original developer is keeping the weights closed at launch, while promising Apache 2.0 in short order.
Our take
Inkling 2 Small represents the first real, full-scale test of the “safe path to open weights” outlined by Thinking Machines Lab (founded by Mira Murati): launch closed, run evaluations, then open-source. The level of training detail (NVFP4, joint RL, the network bypass incident) is rare for a lab of this size and lends credibility to the narrative—yet the Apache 2.0 release is not here yet, and benchmark tables largely come from the developer or voting-based leaderboards. For teams already fine-tuning on Inkling-Small (such as Cognition), the practical question is clear: wait for the promised open release, or begin fine-tuning immediately on Tinker at known prices. Worth watching next week: whether the weights actually drop, and under what exact license.
Sources
- Thinking Machines Lab — Introducing Inkling 2 Small (November 5, 2026) Primary source
- Thinking Machines Lab — Inkling family page Primary source
- Thinking Machines (X) — Inkling 2 Small announcement Primary source
- Tinker — Models & Pricing (Inkling 2 Small) Primary source
- Cognition — Introducing SWE-2 Flash (November 4, 2026) Secondary source — Context: coding model built on Inkling-Small, weights withheld
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
