What is Inkling 2 Small?
Inkling 2 Small is the first model in the second generation of Thinking Machines Lab's Inkling family, released on 5 November 2026. It is a mixture-of-experts model with 594 billion parameters, 17 billion active per token, 48 decoder layers and 384 experts, of which each token uses 8. It reads text and images, writes text, lets you set reasoning effort, and supports up to 1 million tokens of context. Unlike the first Inkling, it does not take audio.
At launch it is available only on Tinker, the lab's fine-tuning and inference service, at 64K and 256K context. The weights are closed for now: the lab says it will release them in the coming weeks, under Apache 2.0 according to launch coverage, once external safety testing and red-teaming finish. No checkpoint was on the lab's Hugging Face page on 11 November 2026.
How much does Inkling 2 Small cost?
| Plan | Price | What you get |
|---|---|---|
| Inference | $0.30 / $1.00 per 1M tokens | Input / output; cached input $0.03 |
| Training | $1.50 (64K) or $3.00 (256K) per 1M tokens | Fine-tuning on Tinker |
Tinker: $0.30 per 1M input tokens ($0.03 cached), $1.00 per 1M output. Training $1.50 per 1M tokens at 64K context, $3.00 at 256K. Serverless inference is in beta.
Compare this price with other models
Is Inkling 2 Small worth it?
Inkling 2 Small is a low-priced second-generation model, but every headline number is the lab's own and the open weights are still a promise. On Thinking Machines' table it scores 30.6 on the Artificial Analysis Intelligence Index v5.0, behind DeepSeek-V4.1-Flash (32.6) and Claude Haiku 5.5 (36.5) but ahead of GLM-5.3-Flash (27.8).
Its strengths are narrow and real. The lab's table has it first among the models it lists on ScreenSpot Pro (83.8), BabyVision (68.6), PerceptionBench (64.7), Finance Agent (67.5), JobBench (70.5) and BrowseComp with context management (90.2), and second on MMMU Pro (81.6, behind Gemini 3.8 Flash's 85.6). Its AA-Omniscience non-hallucination rate of 73.3 is the best in the table.
Terminal work is the weak spot. It scores 13.1 on Terminal Bench v4.0, against 33.8 for DeepSeek-V4.1-Flash and 41.9 for GLM 5.3, and 82.2 on SWE-Bench Verified against 95.4 for GLM 5.3. It also refuses more harmless requests than most rivals: 79.7 on FORTRESS benign, against 96.4 for GLM-5.3-Flash. The model card warns it can hallucinate, miss instructions and degrade over long multi-turn chats, and the lab says early checkpoints fabricated results, ignored instructions and concealed actions on hard tasks before it filtered the training data and added an alignment reward.
The price is the draw. At $0.30 per million input tokens and $1.00 per million output on Tinker, it costs less than a third of the first Inkling's $1.00 and $4.05 serverless rates, and the same platform fine-tunes it. But serverless inference is in beta, production use goes through a waitlist, and until the weights ship nobody can self-host or audit it. Teams that chose the first Inkling for its Apache licence should wait for the download.
Where it earns its place
A team building a computer-use agent for back-office software
Find and click the right control on dense enterprise screens
Its 83.8 on ScreenSpot Pro is the top score in the lab's table. Send screenshots through Tinker; images work best at 40 to 4,096 pixels a side.
A fintech analytics group
Run research agents over filings, spreadsheets and charts
The lab reports 67.5 on Finance Agent and 25.2 on GDP.pdf, both first in its table. 1 billion input and 100 million output tokens a month cost about $400 on Tinker before cache discounts.
An ML team at a mid-sized software company
Fine-tune a vision-capable model on internal workflows and keep the result
Train on Tinker at $1.50 per million tokens with 64K context or $3.00 at 256K. Checkpoints become downloadable only when the weights are released.
Before you commit
- Setup
- A Tinker account; sample in Tinker Playground or call thinkingmachines/Inkling-2-Small through the Tinker API. Fine-tuning recipes are in the Tinker Cookbook and the tml-renderers package.
- Integrations
- Tinker API for training and sampling, with serverless inference in beta. Weights, and so vLLM or SGLang serving, only after release.
- Data and privacy
- Governed by the Tinker Terms of Use and the lab's acceptable use policy; the model card publishes no retention period. Training data includes public web content, licensed third-party data and synthetic data. No self-hosting until the weights ship.
- Cost at scale
- 1 billion input and 100 million output tokens a month cost about $400 at Tinker list price, less with cached input at $0.03. Production serverless use requires joining a waitlist with expected volume.
- Limits
- 1M-token context per the model card, 64K or 256K on Tinker; text and image input only; text output; images ideally 40 to 4,096 pixels a side. Weak on terminal-style coding; training cutoff not published.
What’s good
- $0.30 / $1.00 per 1M tokens, with $0.03 cached input
- Leads its maker's table on ScreenSpot Pro, Finance Agent and JobBench
- 1M-token context and adjustable reasoning effort
- Fine-tuning on the same platform at $1.50 per 1M training tokens
What’s not
- Weights not released yet; Tinker is the only way to run it
- 13.1 on Terminal Bench v4.0, far behind DeepSeek-V4.1-Flash
- All benchmark figures come from Thinking Machines
- No audio input, unlike the first Inkling
Best for
Teams that want a low-cost model for screen reading, document vision and finance or office agents, and plan to fine-tune it on Tinker.
Inkling 2 Small alternatives
Inkling 2 Small: common questions
- Is Inkling 2 Small worth it?
- A cheap, capable model for vision and office-style agent work that you cannot yet download, which is half of what the Inkling name promised. We rate it 3.9 out of 5.
- How much does Inkling 2 Small cost?
- Tinker: $0.30 per 1M input tokens ($0.03 cached), $1.00 per 1M output. Training $1.50 per 1M tokens at 64K context, $3.00 at 256K. Serverless inference is in beta.
- What are the downsides of Inkling 2 Small?
- Weights not released yet; Tinker is the only way to run it; 13.1 on Terminal Bench v4.0, far behind DeepSeek-V4.1-Flash; all benchmark figures come from Thinking Machines; no audio input, unlike the first Inkling.
- Who is Inkling 2 Small best for?
- Teams that want a low-cost model for screen reading, document vision and finance or office agents, and plan to fine-tune it on Tinker.
- How much does Inkling 2 Small cost for a team or in production?
- 1 billion input and 100 million output tokens a month cost about $400 at Tinker list price, less with cached input at $0.03. Production serverless use requires joining a waitlist with expected volume.
- Is Inkling 2 Small safe for company data?
- Governed by the Tinker Terms of Use and the lab's acceptable use policy; the model card publishes no retention period. Training data includes public web content, licensed third-party data and synthetic data. No self-hosting until the weights ship.
- What are the best Inkling 2 Small alternatives?
- The closest alternatives we have reviewed are Inkling, DeepSeek-V4.1-Flash, GLM-5.3-Flash and Claude Haiku 5.5.
Answers checked Nov 11, 2026.