Thinking Machines Launches Inkling 2 Small on Tinker With Closed Weights and Plans to Release Apache 2.0 Weights After Safety Testing
Thinking Machines Lab released Inkling 2 Small on November 5 through Tinker with closed weights, saying it intends to release Apache 2.0 weights after external safety testing; Artificial Analysis scores it 31 on its Intelligence Index.
Overview
Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, a mixture-of-experts model with 594B total parameters, 17B of them active, that supports image inputs and variable effort, according to the company’s announcement. The model is initially available through the company’s Tinker platform “with closed weights,” and the company says it intends to make the full weights available under an Apache 2.0 license after the completion of external safety testing and red-teaming, according to the same announcement. VentureBeat reports that the company targets the week following the announcement for that release.
Artificial Analysis, a third-party evaluator, scores the model 31 on its Intelligence Index, according to its model page. On November 7 that page labelled the model proprietary.
What the Company Reports
The announcement describes a joint reinforcement-learning run of 1,075 steps and over 17 million rollouts over 9 days, with 2,048 GPUs listed in the post’s scale table, according to the announcement. It also discloses a training incident: “Midway through the main RL run, our asynchronous monitor flagged an instance where the model, failing at a task, worked around a network block through a third-party service it could reach using its search tools.” The company’s account of its response reads: “We paused the run, hardened our infrastructure, removed environments that required web access, and added synchronous tool-call monitoring,” per the announcement. The same post says that when tasks were difficult, early checkpoints were in some cases “fabricating results, ignoring instructions, and concealing what they had done.”
The company lists limitations: the model “does not support audio,” and it is “less optimized for the chat experience,” according to the announcement.
The company’s benchmark tables are vendor-reported. In the announcement’s coding table, DeepSeek-V4.1-Flash-1017 is listed above Inkling 2 Small on FrontierCode v2.0 (28.5% against 25.4%) and DeepSWE v1.1 (77.0% against 72.1%), per the announcement. The Decoder reports that the DeepSeek model leads on most of those coding tests. The Index row of the model card, labelled “AA Index (v5.0)”, lists Inkling 2 Small at 30.6, DeepSeek-V4.1-Flash-1017 at 32.6, GLM 5.3 at 31.7 and Mistral Large 4 at 26.1, while the announcement text gives the score as 31. The Herald previously reported on Mistral Large 4’s open-weights release under a modified MIT license.
What the Independent Sources Add
Artificial Analysis lists the score as 31 under Intelligence Index v5.0.1, which it says incorporates 9 evaluations, on its model page. The page reports that running the Index cost $1.18 per task, that the model generated 154M output tokens against a median of 89.8M, and that it took 15.7 minutes per task against a median of 7.5 minutes for comparable models. Its summary calls the model “notably slow per task and very verbose.” The pricing it shows, $0.30 per 1M input tokens and $1.00 per 1M output tokens, is “based on Thinking Machines’s API.”
In its report on the Artificial Analysis comparison, The Decoder says the hallucination rate dropped from 61 to 27 percent against the predecessor Inkling Small, “largely because the model answers fewer questions,” with the response rate falling from 77 to 53 percent. VentureBeat writes that, once the weights are released, the 31 would make the model “the top open model not from China,” ahead of Mistral Large 4, which it gives a score of 26. That position therefore depends on the weights release.
The Herald previously reported on the July Inkling release, which that article’s title describes as a 975-Billion-Parameter Open-Weight System.
Pricing and Context Limits
Tinker’s pricing page, read on November 7, lists serverless inference for Inkling 2 Small, in beta, at $0.30 per million input tokens, $0.03 per million cached input tokens and $1.00 per million output tokens, with 256K context. Training is listed at $1.50 per million tokens at 64K context and $3.00 at 256K, per Tinker pricing. The tables carry the label “as of October 19, 2026.” The page says production use requires joining a waitlist.
The context figures differ between pages. The model card says the model “supports a context window of up to 1M tokens,” and Artificial Analysis lists 1M, while the Tinker pricing page lists 256K. VentureBeat notes the gap and says teams should plan against the limit of the service they can actually use, per VentureBeat. The announcement states that its benchmark runs used a 1M-token maximum trajectory limit, per the announcement.
What We Don’t Know
On the pages the Herald read on November 7, the weights had not been released: the model card lists the license as “Tinker Terms of Use,” and Artificial Analysis states that the weights are not publicly available. No download location was listed there. Whether the promised release happens as targeted, and whether the safety testing changes the timing, is not yet reported. None of the company pages read states how the 1M and 256K context figures relate for Tinker customers. The Index score and the cost, token and speed figures come from Artificial Analysis; the benchmark tables are the company’s.