What happened
Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, describing it as the second generation of its model family. The model is a mixture-of-experts system with 594 billion total parameters and 17 billion active parameters. It accepts image inputs and supports variable effort. The announcement says it will initially be available through Tinker with closed weights, following the company’s previously outlined approach to open-weight releases. Users can try it in Tinker Playground and use it for fine-tuning and sampling on Tinker. Checkpoints are to become downloadable when the weights are released. The announcement does not state a price.
The company reports a score of 31 on the Artificial Analysis Intelligence Index, which it says substantially exceeds its earlier models. It attributes the gains to changes across training, including more efficient pre-training and a large unified reinforcement-learning run. Thinking Machines says Inkling 2 Small was trained with RL on hundreds of thousands of long-horizon coding tasks, across multiple agent harnesses and context-compaction strategies. It says its training process moved beyond unit-test rewards alone: some tasks included an agentic grader assessing how the model approached a problem, not merely whether code passed tests. The company also tested early checkpoints in OpenCode under the name Fledge-Alpha.
The release highlights visual and agentic work, including building and refining websites, games, and 3D designs through visual feedback. Its examples include a procedurally generated city simulator made with Three.js, a Python PC emulator that boots a Linux kernel, and a Blender scene of origami cranes. These are demonstrations selected by the publisher, not a general evaluation of reliability across user projects.
For coding, the company lists Inkling 2 Small at 86.3% on SWEBench Pro v2 (Hard), 72.1% on DeepSWE v1.1, and 25.4% on FrontierCode v2.0. Those figures place it behind another listed model on each of those rows. In the separate PMPP-Hard GPU-kernel optimization evaluation, the release reports a mean of 54.25 tasks solved out of 69 over four attempts and a ten-hour budget. The company says models used maximum available reasoning settings, a shared harness and compaction policy, and independent audits of submitted solutions.
Weights are not yet open. Thinking Machines says it intends to publish them under Apache 2.0 the following week, contingent on external safety testing and red-teaming. That makes open access a stated plan rather than a present capability; the announcement gives no exact publication date or details of the testing process.
What the archive adds
Inkling-Small opens a 12B-active multimodal reasoning model covered the earlier model generation, with 276 billion total parameters, 12 billion active parameters, and a one-million-token context window. This is a new release in a known model family, not a restatement: the new announcement gives a different parameter scale, hosted access, current evaluations, and a prospective weight-license plan. The archive also covered Cognition's SWE-2 Flash, built on Inkling-Small, a downstream use of the previous generation.
Trend signal
This release supports agent economics moves to completed work only weakly: the company emphasizes long-horizon coding tasks, task solving over ten hours, and evaluation beyond unit-test rewards. However, the source supplies no cost-per-completed-task measure, so it does not establish an economic advantage. The evidence is about training and benchmark framing, not production outcomes.
Why now
Thinking Machines ties the current closed-weight availability to a staged release process: Tinker access first, then a planned Apache 2.0 release after external safety testing and red-teaming. The announcement says that testing is still to be completed, but provides no further rationale for the timing or a precise date for the weights.
Connections to other events
The earlier Telegram post on August 17 described Inkling-Small as a 276-billion-parameter MoE with 12 billion active parameters and a one-million-token context, with full weights already available. The current release changes the balance: Inkling 2 Small is much larger in total parameters, but its weights remain closed while the company completes its stated review. Separately, Cognition’s SWE-2 Flash was described in the archive as a coding model built on the earlier Inkling-Small; that shows a downstream model use, not evidence that the new generation is already used in another product.
Why it matters
For teams building coding or visual agents, the release offers a hosted way to test a new model in a real agent loop before deciding whether its benchmark gains matter for their own tasks. The training description is notable because it emphasizes long-horizon tasks, multiple harnesses, context-compaction strategies, and grading of the agent’s approach. Those choices are relevant to teams whose agents must recover from intermediate mistakes rather than simply produce a correct one-shot answer.
The constraints are equally consequential. Tinker is the available route today, while the planned Apache 2.0 weights are not yet available. Teams that require self-hosting, independent auditing, or control over model updates cannot treat the license plan as fulfilled access. The benchmark table also shows uneven standing across tasks, so a headline score or selected demonstration is not enough to choose the model. A team should compare it against its incumbent on representative tasks, with its own tools, review process, and latency or cost requirements before changing production routing.
Limits and uncertainty
The evidence here is the company’s announcement and its reported evaluations. It does not establish independent replication of the Artificial Analysis Intelligence Index score or the coding results, nor does it show performance on a particular team’s repository, visual workflow, or agent harness. The demos are selected examples. The post provides no pricing, latency, throughput, context-window limit, or production availability commitments for Tinker. The Apache 2.0 release is prospective and conditional on external safety testing and red-teaming; the source does not name reviewers, publish test results, or give an exact date. It also does not establish that a planned open-weight release will match the hosted system in behavior.
Reader takeaway
Inkling 2 Small is testable on Tinker now and is positioned for long-horizon coding and visual-agent work, with promising but publisher-reported benchmark results. Treat the Apache 2.0 weights as a pending milestone, not current access. Evaluate the hosted model against your own tasks and wait for the actual weights and safety review before making decisions that depend on self-hosting or license terms.