Inkling 2 Small: Open Weights, Tinker and API Costs

For self-hosting, use the official Hugging Face weights and the matching deployment instructions. The weight-download section below covers formats and multi-GPU requirements.

Playground is the shortest route to a first conversation. For an application, Tinker offers serverless inference in beta. Its current documentation asks production users to join a waitlist with their model, volume and latency requirements. Establish that access before depending on it for a production service.

Sampling and fine-tuning use the Tinker SDK. Keep the route and exact model ID with your usage records: a training client, a sampling client and the serverless endpoint have different billing operations.

Checked November 11, 2026, Tinker lists USD rates per million tokens: $0.30 for fresh input or prefill, $0.03 for cache hits and $1 for sampled output. Training costs $1.50 at 64K context or $3 at 256K. Forward-only calls through a training client use the training rate. Checkpoint storage is separately listed at $0.10 per GB per month.

Open Playground, sign in and select Inkling 2 Small. Supply a short document and ask for a bounded result, such as the payment dates and amounts together with their supporting quotations. Check the quotations against the document. For an image task, use a clear chart and compare the extracted numbers with its labels.

For programmatic access, install the current Tinker SDK and follow its account authentication instructions. The exact model IDs above distinguish the supported configurations. Use the Cookbook sampling workflow with its documented model selection and inspect returned usage before running a batch. Tinker’s compatible-API guide describes sampler checkpoint paths; keep those distinct from serverless model IDs.

For coding, the model needs an agent client that can expose files, execute tools and return results. Start with one repository task and an existing test. Record the complete session, including reasoning, failed edits and retries, when comparing it with another model.

Thinking Machines’ own time-series demonstration has Inkling 2 Small train Qwen3.5-4B through Tinker. The replay ends after 4 hours 17 minutes with $426 spent from a $500 budget. It also describes reducing repeated evaluation-prefill work and stopping an RL run after its score declined. This is a publisher-run workflow, with training and evaluation spend, rather than a price for one Inkling response.

For an autonomous experiment, set separate limits on model calls, training, evaluation and elapsed time. Keep the best held-out result and the complete bill together; low output-token prices alone do not determine the cost of improving another model.

The model supports image and text input and produces text. Thinking Machines reports improvements in agentic coding, reasoning and fine-tuning compared with its first generation. It also identifies weaker chat presentation and tool selection in repositories containing many skills. Audio input is unsupported.

ARC Prize’s verified leaderboard records Inkling 2 Small at 96.5% on the ARC-AGI-1 semi-private set and 81.7% on the ARC-AGI-2 semi-private set at XHigh effort. In ARC Prize’s November 10, 2026 leaderboard snapshot, the results exceeded those of every model it had evaluated whose weights were then open; Inkling’s weights were still pending release at that time. The scores describe ARC-style visual reasoning, not general coding-agent performance.

The model card describes a context window up to one million tokens, while the current priced Tinker configurations above expose 64K or 256K. Size requests for the actual configuration you use. The model card’s maximum is not a one-million-token allowance on every endpoint.

Adam Holter records eight early tasks using Tinker Playground or the native SDK at medium effort. His examples include an SVG request answered with ASCII, website revisions that improved layout, and a small browser game that ran through two waves. Some comparisons use different tools or token caps, so the report is more useful for choosing test cases than ranking models.

For your evaluation, keep the client, instructions, skill files, effort and output budget consistent. Check the generated artifact as well as the response: a claim that a page or game works is easier to assess when you can open it and repeat the intended actions.

The full weights are now available from Thinking Machines on Hugging Face under Apache 2.0. The model card recommends the native NVFP4 checkpoint; MXFP8 is also available. Tinker remains a separate hosted inference and training service with its own model IDs and prices.

The official model card lists at least 500GB aggregate VRAM for NVFP4, with examples of two NVIDIA B300 or four H200 GPUs. Its MXFP8 checkpoint calls for at least 800GB aggregate VRAM, with examples of four B300 or eight H200 GPUs. These are multi-GPU cluster configurations, not desktop requirements; the 17B active count does not represent the stored weight size.

The model card documents vLLM, SGLang, TokenSpeed and Hugging Face deployment frameworks. Budget for the selected runtime, storage, utilization and context cache in addition to accelerator time.

For a smaller local configuration, compare Qwen 27B; for another server-scale checkpoint, see Kimi K3.1 Pro and its deployment conditions.