Key takeaways
- Thinking Machines Lab announced Inkling 2 Small on November 5, 2026. OpenRouter’s free-route record has a November 9 UTC creation timestamp; that is not a new vendor launch date.
- The exact listed API ID is thinkingmachines/inkling-2-small:free, with $0 input and $0 output token prices, a 1,048,576-token context and a 131,072-token maximum completion.
- Thinking Machines describes a mixture-of-experts model with 594B total parameters and 17B active, image inputs and variable reasoning effort. The model does not support audio.
- The exact free-route endpoint lists Thinking Machines as provider and nvfp4 quantization. Tool support is listed, but explicit tool_choice controls are not advertised.
- The November 5 announcement promises Apache 2.0 weights after external testing. A promise is not a verified weight release; do not plan self-hosting from this briefing.
- AccessAllGPT did not test the model because the API key was unavailable. Use the free route for bounded evaluations, not as evidence of production capacity or reliability.
What changed: a free route for the November 5 model
Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, initially through Tinker for inference and fine-tuning. OpenRouter now lists thinkingmachines/inkling-2-small:free at $0 per million input tokens and $0 per million output tokens. The registry record was created on November 9 UTC and was directly retrieved for this article on November 10. That gives builders another listed access path; it does not change the model’s original announcement date.
OpenRouter advertises a 1,048,576-token context and a 131,072-token maximum completion, accepting text and images and returning text. Its exact free-route provider API lists a Thinking Machines endpoint with nvfp4 quantization. These are observed registry facts, not results from an AccessAllGPT inference call. Context capacity is not evidence that a model accurately uses every part of a million-token input.
The practical opportunity is inexpensive evaluation of an agent-oriented model without first committing to a fine-tuning workflow. Free token pricing does not establish unlimited requests, reserved capacity, data-handling terms or a service-level guarantee. Check the current route and account conditions before sending anything beyond public or synthetic test material.
Small means active parameters, not a small download
Thinking Machines describes Inkling 2 Small as a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. The active count describes computation within a much larger model; it should not be read as the storage or memory requirement of a conventional 17B dense model. This briefing does not estimate self-hosting hardware.
The announcement emphasizes long-horizon coding, visual feedback and knowledge work. Thinking Machines says it trained on hundreds of thousands of long-horizon coding tasks across agent harnesses and context-compaction strategies, using agentic graders as well as unit-test rewards. It also identifies Fledge-Alpha on OpenCode as an early checkpoint. Those are vendor descriptions of training and testing, not an independent audit.
The launch post presents coding evaluations and demonstrations, but we did not reproduce them. Readers should not treat a polished simulator demo or a published benchmark table as a prediction of correctness in their own repository. A fixed issue set with executable acceptance tests remains more useful than switching on the strength of the launch material alone.
Reasoning and tools need an integration check
OpenRouter marks reasoning as mandatory and enabled by default. It lists max, high, medium, low and minimal effort, with high as the default. Record effort explicitly in a comparison: a change in reasoning settings can change latency and output volume even when the advertised token price is zero.
The registry lists tools among the supported parameters, but not tool_choice. The free-route endpoint’s supports_tool_choice fields are all false. Do not assume an application that forces a particular function, or requires a tool call on every turn, can migrate unchanged. Test the exact request shape and inspect actual calls before wiring the model into a privileged agent.
The distinction between the exact ID and a related model path matters. The endpoint for thinkingmachines/inkling-2-small:free returned a provider listing during our review; the canonical-slug endpoint without the free suffix returned an empty endpoint array. Use the exact advertised free ID rather than inferring that every similarly named route is available.
The vendor flags chat and tool-selection weaknesses
Thinking Machines says Inkling 2 Small is not its most succinct communicator and does not always choose the best tools, especially in codebases with many skills. It also reports weaker organization and presentation in chat, including overuse of slashes. These are useful vendor-disclosed limitations for teams deciding whether to evaluate it as an autonomous coding worker or a customer-facing assistant.
The model supports image input but not audio, according to the announcement. OpenRouter’s text-and-image input listing is consistent with that narrower interface. Do not infer speech, video or image generation support from the word multimodal or from demos where an agent writes code that produces visual output.
For an initial trial, use a small tool set, public fixtures and reversible tasks. Measure whether the model picks the right tool, follows file boundaries and verifies changes. Keep customer data and secrets out of the trial until the exact provider’s retention and access policies have been reviewed.
Open weights remain a separate verification step
The November 5 announcement says Inkling 2 Small initially has closed weights on Tinker and states an intention to release full weights under Apache 2.0 the following week, after external safety testing and red-teaming. It says downloadable fine-tuning checkpoints will accompany that release. This article does not establish whether those artifacts have subsequently shipped.
A free managed API route is not an open-weight release. We did not verify a downloadable repository, weight files, license file, hashes or serving recipe for Inkling 2 Small. If your purchase decision depends on self-hosting or portable fine-tuned checkpoints, wait for those actual artifacts and their terms rather than treating the roadmap as delivery.
What to do next
If you already use OpenRouter, check the exact free route and run a bounded comparison against your current model using public code or synthetic documents. Include an image-input task, a small tool-selection task and a long-context task with a known answer. Record effort, token limits, completion status, corrections and end-to-end time.
AccessAllGPT did not run that comparison: OPENROUTER_API_KEY was not set in this environment. This is a source-backed briefing, not a Hands-on review, and it contains no generated UI screenshots or measured performance claims. The next step is to evaluate access and quality, not to promote a zero-price listing into a production dependency.
A bounded free-route trial
Use these checks before deciding that the new access path fits an existing agent.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Request thinkingmachines/inkling-2-small:free; confirm provider, current quota and data terms rather than guessing from the model family name.
Fix reasoning effort and output limits. Test image input and tool calls; do not rely on unsupported explicit tool_choice controls.
Run public tasks with known acceptance criteria. Log tool mistakes, unsupported claims, corrections and latency, including failed requests.
Keep the existing model as fallback. Treat self-hosting as unapproved until actual weight and license artifacts are verified.
Primary sources
Browse the publication-wide evidence index →
- Introducing Inkling 2 SmallThinking Machines Lab · Reviewed: November 5 date; introduction and initial access; agentic coding; limitations; try Inkling 2 Small · Retrieved · Supports: Vendor announcement date, mixture-of-experts parameter counts, image support, Tinker inference and fine-tuning, coding training description, disclosed limitations and conditional future Apache 2.0 weight-release plan.
- Inkling 2 Small (free) model pageOpenRouter · Reviewed: Exact model route and public model overview; registry details cross-checked against the Models API · Retrieved · Supports: Public discovery page for the free OpenRouter route. Listing presence is not proof of a successful account-level inference call.
- OpenRouter Models API: Inkling 2 Small free recordOpenRouter · Reviewed: thinkingmachines/inkling-2-small:free identity, created timestamp, context, completion limit, modalities, pricing, reasoning and supported parameters · Retrieved · Supports: November 9 UTC registry creation, zero input/output token rates, 1,048,576-token context, 131,072-token completion maximum, mandatory reasoning and high default effort.
- Exact free-route provider endpointOpenRouter · Reviewed: Provider endpoint name, model_id, pricing, quantization, context, supported_parameters and supports_tool_choice · Retrieved · Supports: Directly observed Thinking Machines provider listing, nvfp4 quantization and zero token rates; tools are listed but explicit tool-choice capabilities are false.
Limitations
We did not invoke Inkling 2 Small because OPENROUTER_API_KEY was not set, run hands-on UI tests, measure latency, test context accuracy, inspect training data, reproduce vendor benchmarks or verify weight availability, model-card details, fine-tuning export, quotas or retention policies. Registry fields and provider availability can change. The original November 5 announcement and November 9 registry timestamp describe different events. Biological applications are excluded from this unattended briefing.
Disclosures
AccessAllGPT received no payment, model credits or vendor briefing for this article. Thinking Machines Lab and OpenRouter did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with these vendors. Publication-wide relationships are listed on the disclosures page.
