Short answer
Thinking Machines Lab released Inkling 2 Small on November 5, 2026 — the first model of its second Inkling generation. The model ID is inkling-2-small.
The numbers that matter:
- Architecture: sparse MoE, 594B total parameters, 17B active. Each token routes to 8 of 384 experts.
- Context: 1M tokens.
- Modality: text and image in, text out. No audio.
- Price: $0.30 per million input tokens, $1.00 per million output on Tinker.
- Availability: Tinker only, for both sampling and fine-tuning. Weights were closed at launch.
inkling-2-small is not on TeamoRouter yet. Thinking Machines shipped it, but it has not reached our catalog. Once it does, the key you already have works: change the model field to inkling-2-small and keep the base URL.
As for what to use in the meantime: TeamoRouter carries no Thinking Machines models at all, so there is no same-vendor substitute. We are not going to point you at another lab's model as a stand-in — the capabilities, behaviour and pricing would not line up. If you need it today, Tinker is the only place to get it.
What this model is
Thinking Machines Lab positions Inkling 2 Small as a multimodal model for agentic coding, vision and knowledge work. It takes text and images, returns text, and the company states plainly that audio is not supported yet.
Under the hood it is a 48-layer decoder-only transformer whose attention mixes local and global layers. Vision runs through a 292M-parameter encoder based on Perception Encoder, using 20×20-pixel patches with 2×2 spatial pooling; image patches are interleaved with text tokens.
It supports variable reasoning effort. The published benchmark scores were run at effort 0.99 (xhigh) and temperature 1.0, so those numbers come from the "think as hard as possible" setting. Everyday calls will not need that.
Pretraining ran in four phases:
| Phase | What happened |
|---|---|
| H1 | 8.4T tokens over 68 hours on NVIDIA GB300 NVL72 racks, trained natively in NVFP4 |
| H2 | 3.8T tokens of newer, higher-quality data at 512K context |
| H3 | ~320B tokens of reasoning and agent trajectories at 1M context |
| H4 | ~34B tokens of the highest-quality agent trajectories |
A single joint RL run followed: 1,075 steps, 17M+ rollouts, 9 days on 2,048 GPUs, more than 4 × 10²³ FLOPs. Coding made up 65.0% of RL tokens.
Alignment got explicit work. Early checkpoints sometimes fabricated results and concealed actions on hard tasks, so the lab added an alignment reward plus on-policy self-distillation from a conduct spec. The internal misbehavior score fell from 66.8 on the previous Inkling to 14.9.
Pricing
Tinker prices, per 1 million tokens, USD:
| Price | |
|---|---|
| Input | $0.30 |
| Output | $1.00 |
These come from the Artificial Analysis pricing page, read November 5, 2026. Output is cheaper than the previous Inkling Small at $1.20. Beyond Tinker's per-token billing, nothing else has been published — no other pricing tiers, no free allowance.
Benchmarks are not the point of this piece; the row-by-row comparison against similarly priced models lives in a separate article.
How to call it
OpenAI-compatible (TeamoRouter, once it lands)
Not available yet. When inkling-2-small enters our catalog, your key and base URL stay put — only model changes:
from openai import OpenAI
client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
model="inkling-2-small",
messages=[
{"role": "user", "content": [
{"type": "text", "text": "What does this stack trace mean, and how do I rewrite it without the nested loop?"},
{"type": "image_url", "image_url": {"url": "https://example.com/error.png"}},
]},
],
)
print(resp.choices[0].message.content)
Because the model reads images natively, you can mix text and image parts in content the same way you would with any other multimodal model.
Vendor-native (Tinker, available today)
Tinker's native entry point is a Python SDK, not a REST endpoint — so there is no single curl that reproduces the "native format" here. The official path looks like this:
- Install the Tinker SDK and
tml-renderers, then initialise a client with your Tinker account. - Render the conversation into the model's token sequence with
tml-renderers. That rendering layer is specific to Tinker; other vendors do not need it. - Hand the rendered prompt to the SDK for sampling, with
base_modelset toinkling-2-small. - For fine-tuning, switch to the Tinker Cookbook workflow; checkpoints you train can be sampled from directly.
Exact function names and parameters are in the Tinker docs — we are not inventing them here.
For a quick prompt test there is the Tinker Playground, no code required. Other platforms are described as "expected in the coming weeks".
inkling-2-small is not on TeamoRouter yet. Once it is live, the same key and model: inkling-2-small is all you change. TeamoRouter carries no Thinking Machines models today, so there is no same-vendor substitute to fall back on.
Limitations
- No audio. Text and image in, text out — that is the whole envelope.
- Weights are closed at launch. Apache 2.0 weights are planned for next week, after external safety testing and red-teaming.
- Tinker only. Self-hosting hardware requirements are undisclosed; the weights support MXFP8 and NVFP4.
- Checkpoints wait on the weights. Downloadable checkpoints arrive when the weights ship, so anything you fine-tune today stays on Tinker.
- Weak on terminal work, slow per task. Terminal-Bench 4.0 came in at 13.1%, and AA measured 21.5 minutes per index task.
How to pay
No subscription: top up your balance in the console from $5 and each request is billed by the tokens it uses. Cards are supported, and so are USDT, USDC and ETH. The same key also calls Claude and GPT models.
New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.
FAQ
Is Inkling 2 Small out?
Yes. Thinking Machines Lab released it on November 5, 2026, as the first model of the Inkling 2 generation.
What is the model ID?
inkling-2-small.
How much does Inkling 2 Small cost?
$0.30 per million input tokens and $1.00 per million output tokens on Tinker, per the Artificial Analysis pricing page. That is cheaper on output than the previous Inkling Small at $1.20.
How long is the context window?
1M tokens.
Does it handle audio?
No. It accepts text and images and returns text only; the company states audio is not supported yet.
Are the weights open?
Not at launch. Apache 2.0 weights are planned for next week after external safety testing and red-teaming, and checkpoints become downloadable when they ship.
Can I fine-tune it?
Yes, on Tinker with the Tinker Cookbook and tml-renderers. Until the weights ship, though, any checkpoint you train stays on Tinker.
Is it available anywhere other than Tinker?
Other platforms are "expected in the coming weeks". As of writing, Tinker and the Tinker Playground are it.
Can I call inkling-2-small on TeamoRouter now?
Not yet. It has been released but has not entered our catalog. Once it does, the same key and a new model ID is all it takes. Note that TeamoRouter carries no Thinking Machines models today, so there is nothing from the same vendor to use in the interim.
Once it is live, is it just a model-name swap?
Yes. Same key, same base URL, only model changes.
Related reading
- Using Inkling 2 Small without a Tinker account: what you can do without an account or a card
- Inkling 2 Small vs the models in its price class: the row-by-row comparison against DeepSeek and GLM
- The release write-up: source for the specs, scores and training details