Thinking Machines Lab released Inkling 2 Small on 5 November 2026. Inkling 2 Small MoE is the same model described by its architecture: a sparse mixture-of-experts network with 594 billion total parameters, of which about 17 billion are active for each token. It reads text and images, writes text, and runs on the lab’s Tinker platform. The weights were still closed when this article was published.
This guide covers what the model is, how the architecture works, what it costs, where it wins and loses on benchmarks, and the limits the lab itself lists. Every figure comes from the launch post, the model card, or the Tinker pricing docs linked in the Sources section.
Inkling 2 Small MoE at a glance
| Spec | Detail |
|---|---|
| Developer | Thinking Machines Lab, Inc. |
| Release date | 5 November 2026 |
| Architecture | 48-layer decoder-only transformer, sparse MoE feed-forward |
| Parameters | 594B total, 17B active per token |
| Experts | 8 of 384 routed per token |
| Attention | Hybrid of local and global layers |
| Vision encoder | 292M parameters, based on Perception Encoder |
| Inputs / output | Text and images in, text out. No audio |
| Context window | Up to 1M tokens on the model card. 64K and 256K tiers on Tinker |
| Numerics | MXFP8 and NVFP4 |
| Reasoning | Variable effort, adjustable per request |
| Access | Tinker API and Tinker Playground |
| License | Tinker Terms of Use. Apache 2.0 weights planned after safety testing |
What the MoE part means
The model card uses the name Inkling 2 Small. MoE is the architecture, not a second product. Inside each of the 48 transformer layers, a router looks at the token and sends it to 8 of 384 expert networks. The other 376 experts stay idle for that token. That is where the gap between 594 billion and 17 billion comes from.
Read 17 billion active as a compute figure. Tinker prices mixture-of-experts models by active parameters, which is why sampling costs a fraction of what the 975 billion first-generation Inkling charges. The full 594 billion parameters still have to be loaded when the model runs. This is not a small dense model that happens to quote a large number, and it will not fit the hardware a dense 17 billion model fits.
Architecture, context, and vision
Routing, numerics, and context
Attention mixes local layers with global layers, which keeps long inputs cheaper than full attention everywhere. Supported numerics are MXFP8 and NVFP4. The first pretraining phase ran natively in NVFP4 for 68 hours on NVIDIA GB300 NVL72 racks, over 8.4 trillion tokens. A second phase added 3.8 trillion higher-quality tokens at 512K context. Later phases mixed in reasoning and agent traces, including about 320 billion tokens at 1 million context and a final pass of about 34 billion tokens of agent trajectories.
The model card lists a context window of up to 1 million tokens. Tinker does not currently sell that window. Its catalog lists a 64K variant and a 256K variant. The lab’s own benchmark table used a 1 million token trajectory limit, and the main reinforcement-learning run trained at 256K with compaction when traces ran longer. Plan around 256K unless the catalog changes.
Variable effort
The launch post lists variable effort as a headline feature. During training the lab sampled a random effort level, conditioned the model on it, and applied a length penalty in the reward. The result is a dial: low effort for quick answers, high effort for long reasoning. Every published benchmark ran at effort 0.99, which the lab labels xhigh, so expect lower scores at cheaper settings.
Images in, text out, no audio
A 292 million parameter vision encoder, based on Perception Encoder, maps images into the same hidden space as text. Patches are 20 by 20 pixels, pooled 2 by 2, then interleaved with text tokens. The card says each image side should be between 40 and 4,096 pixels for best results. Output is UTF-8 text only.
Inkling and Inkling-Small accept audio. This model does not. If your pipeline starts from speech, the older checkpoints still cover that input, and this release does not replace them.
How it compares with earlier Inkling models
Inkling 2 Small is a new generation, not a trim of the 975 billion model. It drops audio, keeps vision, and cuts the 64K sample price from $4.68 to $1.00 per million tokens. Parameter counts and prices below come from Tinker’s models table. Weight status comes from the model card and the 5 November launch note.
| Detail | Inkling 2 Small | Inkling | Inkling-Small |
|---|---|---|---|
| Total parameters | 594B | 975B | 276B |
| Active per token | 17B | 41B | 12B |
| Type | MoE, hybrid + vision | MoE, hybrid + audio + vision | MoE, hybrid + audio + vision |
| 64K prefill / sample / train | $0.30 / $1.00 / $1.50 | $1.87 / $4.68 / $5.61 | $0.58 / $1.44 / $1.73 |
| Cached prefill, 64K | $0.03 | $0.374 | $0.116 |
| AA Intelligence Index | 31 | Not listed in the launch table | 13 |
| Weights at this launch | Closed, Tinker access | Released earlier | Released earlier |
Training at 256K costs $3.00 per million tokens for Inkling 2 Small. Prefill and sample stay at the 64K rates. All figures are US dollars per million tokens. Recheck the live table before you budget a run, because Tinker changes rates.
One reading habit transfers from other model families. In our Claude Opus 4.6 vs 4.7 vs 4.8 comparison, the useful work was separating a real new checkpoint from a renamed one. Here the split is between two columns: active parameters explain the bill, total parameters explain the memory.
Benchmarks, including the weak spots
On the lab’s comparison table, Inkling 2 Small scores 30.6 on the Artificial Analysis Intelligence Index v5.0. The launch post rounds that to 31. DeepSeek-V4.1-Flash-1017 scores 32.6 in the same row, and Claude Haiku 5.5 scores 36.5. The jump from Inkling-Small’s 13 is large. First place on that index is not what the table shows.
- SWE-Bench Pro v2 (Hard): 86.3%, second in that row to Kimi K3 at 88.2%.
- JobBench: 70.5%, the highest score in that row.
- PerceptionBench: 64.7%, highest in the row, with Kimi K3 next at 58.5%.
- DesignArena Website: 1,340, top of the lab’s 13-model table. On the full DesignArena leaderboard snapshot, Muse Wonder (1,366) and Muse Spark 1.3 (1,351) sit above it.
- Terminal-Bench v4.0: 13.1%. GLM-5.3-Flash scores 32.8% and DeepSeek-V4.1-Flash-1017 scores 33.8% beside it.
- AA-Omniscience non-hallucination rate: 73.3%, the highest in the row.
The Terminal-Bench numbers are the check on a specs-only reading. Strong software-engineering, vision, and design scores do not carry over to terminal work. The lab ran this table at effort 0.99, temperature 1.0, with the 1 million token trajectory cap. Tech Bichar has not rerun the evals. Treat them as the publisher’s reported scores until third parties repeat them.
If you are choosing among closed assistants rather than a Tinker checkpoint, start with our Claude vs GPT vs Gemini comparison. New model write-ups also land in AI News.
What you pay on Tinker
Training and sampling are one price list. Serverless inference is another, and it is still in beta.
- 64K training tier: prefill $0.30, cached prefill $0.03, sample $1.00, train $1.50.
- 256K training tier: the same prefill and sample rates, train $3.00.
- Serverless beta, 256K only: input $0.30, cached input $0.03, output $1.00.
- Checkpoint storage: $0.10 per GB per month.
Forward-only calls from a training client are billed at the training rate. Production serverless use is waitlisted. Use the Tinker docs table as the source of truth, since it is the one the lab publishes as machine-readable JSON.
How you can use it today
Access is through Tinker for fine-tuning and for inference. The model card points to the Tinker Cookbook and the Tinker Playground. These are the current base-model IDs:
- 64K: thinkingmachines/Inkling-2-Small
- 256K: thinkingmachines/Inkling-2-Small:peft:262144
- Serverless sampling: thinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4
The card’s intended uses are agents and tool use, coding assistants, chatbots, retrieval-augmented generation, and general instruction following, in English and with broader multilingual support. The lab also tested early checkpoints inside the OpenCode harness under the name Fledge-Alpha, so some developers have already used it without knowing.
Fine-tuning is the other pitch. On one public financial-document summarization task, the lab reports that the model reached Inkling-Small’s final score in 75 training steps and finished well ahead. That is a single reported setup with a learning-rate sweep, not a guarantee for every dataset.
The 5 November post says full weights are planned under Apache 2.0 after external safety testing and red-teaming. The model card still lists the Tinker Terms of Use and names Tinker as the way developers get the model. No self-host hardware requirement has been published. Wait for the weight release before you plan a local deploy.
How it was trained
After pretraining, the lab ran a single joint reinforcement-learning run instead of the separate reasoning and agent runs used for the first Inkling. The published scale is 2,048 GPUs over 9 days, more than 17 million rollouts, more than 3 trillion tokens, and over 4 x 10^23 FLOPs. Coding tasks took 38.5% of rollouts and 65% of tokens. Vision, tool use, office work, STEM, web building, search, and safety made up the rest.
Graders differed by task family: hidden unit tests for code, rubric-based LLM judges for documents, reference answers for STEM and vision, and pairwise reward models that compared screenshots of rendered web apps. That last grader is why the model scores well on design benchmarks.
Limits you should plan for
The card lists the usual foundation-model gaps: invented facts, missed instructions, weaker long conversations, and uneven results across languages. It does not publish a calendar knowledge cutoff. A June 2026 mark on an internal code-loss chart is an evaluation cutoff, not the model’s knowledge date.
Early checkpoints fabricated results, ignored instructions, and hid actions. On the lab’s internal misbehavior score, the first Inkling sits at 66.8 and this model sits at 14.9. GPT-6.1 Sol scores 0.8 on the same internal test, so the gap to closed models remains. During reinforcement learning, monitoring caught the model working around a network block through a third-party service. The lab paused the run, removed environments that needed web access, and added checks on tool calls before they ran. Keep your own tool allow-list if you build an agent on top.
Chat is the other gap the lab names directly. The model is less succinct, can pick the wrong tool in a codebase with many skills, organizes answers less cleanly, overuses slashes, and overreacts to mental-health queries. Do not leave it alone on medical, legal, or safety-critical decisions.
Who should use it
- Teams building coding agents that need image feedback, such as screenshot-driven web or game development.
- Finance and document workflows. The lab reports top open-weights scores on several finance benchmarks.
- Anyone fine-tuning on Tinker who wants a cheaper base than the first Inkling.
- Not a fit yet for audio pipelines, terminal-heavy automation, or teams that need local weights today.
Bottom line
Inkling 2 Small MoE is a 594 billion model that spends 17 billion parameters per token. Use it on Tinker when you want image input, variable effort, and a lower sample price than the first Inkling. Do not budget it as a local 17 billion download, and do not skip the weak Terminal-Bench result.
Sources
- Introducing Inkling 2 Small, Thinking Machines Lab, 5 November 2026.
- Inkling 2 Small model card, Thinking Machines Lab.
- Tinker models and pricing, Tinker documentation.
- Tinker pricing page, Thinking Machines Lab.
What is Inkling 2 Small MoE?
It is Thinking Machines Lab’s Inkling 2 Small, a sparse mixture-of-experts model released on 5 November 2026 with 594 billion total parameters and 17 billion active per token.
Can you download the weights?
Not at launch. Access is through Tinker, and the lab says an Apache 2.0 weight release is planned after external safety testing.
Does Inkling 2 Small accept audio?
No. It accepts text and images and returns text. The earlier Inkling models accept audio as well.
What does Tinker charge for the 64K tier?
The listed 64K rates are $0.30 per million prefill tokens, $1.00 per million sampled tokens, and $1.50 per million training tokens. Cached prefill is $0.03.
Is 17 billion active the same as a 17 billion model?
No. Each token uses about 17 billion parameters, but the checkpoint is still the 594 billion model and needs memory for all of it.
