Inkling 2 Small launches on Tinker, with open weights pending
Thinking Machines Lab introduced Inkling 2 Small on November 5 with inference and fine-tuning through Tinker. The weights remain closed at launch. Developers can evaluate the hosted model now, but should not plan around an already available Apache 2.0 download.
The model’s advertised capacity differs from the context available through Tinker today.
What Inkling 2 Small supports
The model card describes a mixture-of-experts transformer with 594 billion total parameters and 17 billion active. It accepts text and images, produces text, and supports up to one million tokens of context. Its 48-layer architecture routes each token to eight of 384 experts.
Those two parameter counts answer different questions. Active parameters describe computation selected for a token; total parameters describe the larger model. A 17-billion-active label is not a promise that the complete model fits the memory budget of a 17-billion-parameter dense model.
The card currently lists Tinker Terms of Use. Thinking Machines intends to release Apache 2.0 weights after external safety testing and red-teaming, targeting the week after the announcement. Downloadable fine-tuned checkpoints are tied to that release. This remains a plan, not a completed license transition.
Our earlier coverage of the original Inkling family documents its separate July releases. Do not confuse those existing downloads with the new second-generation model.
Tinker context limits and current prices
The current Tinker catalog lists 64K and 256K training variants. Serverless inference is a 256K-context beta, with a waitlist for production use. These service limits are narrower than the model card specification.
| Tinker operation | Price per million tokens |
|---|---|
| Uncached input / prefill | $0.30 |
| Cached input / prefill | $0.03 |
| Output / sampling | $1.00 |
| Training at 64K context | $1.50 |
| Training at 256K context | $3.00 |
These are the documented USD rates checked on November 6. Checkpoint storage adds $0.10 per GB per month. The catalog bills forward-only calls made by training clients at the training rate, so they should not be budgeted as ordinary inference.
Across multiple requests, one million uncached input tokens plus 100,000 output tokens cost $0.40, excluding training, storage and other usage.
Before committing a workload, confirm the exact endpoint, supported sequence length and production availability. A context ceiling, a token price and an application-level latency target are separate constraints.
Reasoning effort is a setting to test
Tinker’s thinking-effort guide uses a continuous value from zero up to, but excluding, one. The default high preset is 0.9; xhigh is 0.99. Higher settings generally encourage more reasoning without guaranteeing a better answer on each sample.
Effort and the output-token cap are independent. Raising effort without enough generation budget can truncate a response. The documentation recommends comparing multiple samples or aggregate accuracy rather than treating one impressive answer as an evaluation. Zero effort also does not enforce a hard prohibition on reasoning.
What to check before deployment
The launch notes acknowledge imperfect tool selection in skill-heavy codebases and less polished chat presentation. Unlike earlier Inkling models, this release does not support audio. ByteForward has not independently tested the model.
The model card also warns about hallucinations, imperfect instruction following and degraded behavior in long conversations. Test your own repository, image inputs, languages and failure cases, then compare quality at a fixed token and time budget.
Illustration by ByteForward, generated with AI using the official Thinking Machines wordmark for editorial identification. The equipment is conceptual; no endorsement is implied.
