If you are evaluating Inkling, the immediate decision is whether you need Tinker’s 256K context option enough to pay roughly twice the listed 64K rates. The model weights are open, but managed sampling and fine-tuning on Tinker are usage-priced. The launch discount also has no published end date, so it should not be treated as a permanent project-cost baseline.
Prices and specifications in this article are as of 2026-07-20.
What Inkling is
Thinking Machines Lab released Inkling on July 15, 2026, according to the official Inkling model card. It is a multimodal Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters. It accepts text, images, and audio and produces text.
The weights are available under the Apache 2.0 license. Thinking Machines says the full model supports up to a 1 million-token context window and distributes downloadable checkpoints through Hugging Face. That does not mean every hosted route exposes 1 million tokens: the managed Tinker options are capped at 64K or 256K.
The official launch post presents Inkling as an open-weights base for customization rather than the strongest general model overall. It is available for fine-tuning through Tinker, while third-party providers offer separate inference APIs. Those providers have their own pricing, which is outside the Tinker rates below.
Thinking Machines Inkling pricing on Tinker
Tinker’s official models and pricing page lists all usage prices in US dollars per million tokens. As of 2026-07-20, it shows:
| Tinker option | Model ID | Context | Prefill | Cached prefill | Sample | Train |
|---|---|---|---|---|---|---|
| Inkling | thinkingmachines/Inkling | 64K | $1.87 | $0.374 | $4.68 | $5.61 |
| Inkling extended context | thinkingmachines/Inkling:peft:262144 | 256K | $3.74 | $0.748 | $9.36 | $11.23 |
Prefill is input processing, sample is output generation, and train covers the forward and backward pass used to compute gradients. Cached prefill receives the documented 80% discount when input tokens hit Tinker’s prompt cache. Checkpoint storage is listed separately at $0.10 per GB per month.
Both Inkling rows currently carry a “Limited-time 50% discount” label. Thinking Machines does not state an end date on the launch post or pricing page. The figures above are the rates displayed while that label is active; they should be rechecked before a production run. We would not double them and call that a confirmed post-promotion price because the company has not published that future rate.
Which context option should a developer choose?
The 64K option is the more conservative starting point for short supervised fine-tuning examples, compact agent traces, or evaluations that fit comfortably within that sequence limit. Its listed prefill, sampling, and training rates are about half those of the 256K route.
The 256K option is relevant when individual training or sampling sequences genuinely exceed 64K—for example, long repository context, extended tool histories, or large multimodal records. It uses the :peft:262144 model ID and carries higher rates across every metered token category. Choosing it for hypothetical future headroom makes each token more expensive even when a run never approaches the larger limit.
Open weights give developers another path: download and operate Inkling outside Tinker. That avoids Tinker’s token charges, but it does not make deployment costless. The model card says the BF16 checkpoint requires at least 2 TB of aggregate VRAM, while the NVFP4 checkpoint requires at least 600 GB. Hardware, orchestration, and engineering costs therefore belong in any self-hosting comparison.
Who should start with each route
Start with Tinker’s 64K option if your immediate examples, agent traces, evaluations, or sampling requests fit comfortably inside that context. It carries the lower listed rate in every token category, so it lets you test Inkling’s managed customization path without paying for unused extended context.
Consider the 256K route when an individual sequence actually crosses the 64K boundary. The reason to choose it is the sequence you need to process now, not the possibility that a future project might be larger. The :peft:262144 model ID and higher prefill, cached prefill, sample, and train rates make that distinction worth recording in your configuration and budget.
Treat self-hosting as a separate operating decision, not the free version of Tinker. Open weights remove Tinker’s managed token charges, while the stated checkpoint requirements move the comparison toward hardware, orchestration, and engineering. A team prepared for at least 2 TB of aggregate VRAM for BF16 or at least 600 GB for NVFP4 is evaluating a different cost structure from a developer buying managed tokens.
A pre-run cost checklist
- Measure the longest individual sequence you intend to use. Compare it with 64K before selecting 256K for headroom.
- Separate the token categories. Estimate prefill, cached prefill, sample, and train usage rather than applying one rate to the whole run.
- Identify which inputs can hit Tinker’s prompt cache. Apply the documented cached-prefill price only to tokens that receive that treatment.
- Include checkpoint storage. The listed $0.10 per GB per month sits outside the per-million-token rows.
- Mark the discount as temporary in the budget. The page shows a limited-time 50% discount but gives no end date or confirmed replacement rate.
- Keep hosted routes separate. Tinker’s 64K and 256K options do not expose the full 1 million-token context described for the model, and third-party inference pricing is outside this table.
- If comparing self-hosting, include the stated checkpoint requirements. Open weights change who operates the model; they do not remove hardware or engineering from the decision.
What would change the recommendation
If your sequence length grows beyond 64K, the extended-context route becomes a current requirement rather than hypothetical insurance. If you shorten examples or tool histories so they fit within 64K, the lower-rate route becomes relevant again. The decision should follow the material you actually process.
A published end date or replacement price for the launch discount would also change the budget. Until Thinking Machines states one, the current discounted rows are verifiable and a doubled future rate is not. Keep those two statements separate.
Self-hosting becomes a stronger comparison only when your team can account for the checkpoint’s aggregate VRAM needs and the associated orchestration and engineering. Conversely, if you want managed sampling and fine-tuning without operating that hardware, Tinker remains the directly priced route described here.
Finally, do not use the model’s 1 million-token ceiling to answer a Tinker question. The managed options in this article stop at 64K and 256K. A provider that exposes a different context or price would require a separate comparison rather than a silent substitution into this one.
Bottom line
Inkling offers open weights plus a managed customization route. For Tinker users, 64K is the lower-cost default and 256K is a specific extended-context purchase, not the model’s full 1M-token ceiling. Budget using the current per-token categories, include checkpoint storage where relevant, and treat the 50% launch discount as temporary until Thinking Machines publishes a firm end date or replacement pricing.