This review is built from what could be checked on November 6, 2026, the day after launch: Thinking Machines Lab's announcement, model card, Inkling page, Tinker documentation, prices and terms; independent results from Artificial Analysis, Vals AI and DesignArena; user reports on Hacker News; reporting by TechCrunch, WIRED, Unite.AI and Crypto Briefing; and the price pages of four rival model makers. Benchmark figures are Thinking Machines' own claims unless another source is named.
What Inkling 2 Small is
Thinking Machines Lab is the company Mira Murati started after leaving her post as OpenAI's chief technology officer; she announced it in February 2025 as chief executive, with OpenAI co-founder John Schulman as chief scientist (TechCrunch). In July 2025 it closed a 2 billion USD seed round led by Andreessen Horowitz at a 12 billion USD valuation, which WIRED called the largest seed round on record (TechCrunch, WIRED). In March 2026 NVIDIA invested again as part of a deal to deploy at least one gigawatt of its Vera Rubin systems for the company (Thinking Machines).
The company sells Tinker, an API for fine-tuning open-weight models (Tinker), and since July 2026 it has released its own models, the Inkling family (Inkling page). Inkling 2 Small is the first of the second generation. It is a mixture of experts with 594 billion parameters, of which 17 billion work on each token: a 48-layer decoder sends every token to 8 of its 384 experts. It reads text and images, writes text, and the model card gives a context window of up to 1 million tokens (model card). Despite the name it is more than twice the size of the first Inkling-Small, and Thinking Machines calls it its most capable model.
| Inkling 2 Small | Inkling | Inkling-Small |
|---|
| Released | November 5, 2026 | July 15, 2026 | July 30, 2026 |
| Parameters | 594B, 17B active | 975B, 41B active | 276B, 12B active |
| Inputs | Text, images | Text, images, audio | Text, images, audio |
| Weights | Closed at launch, Apache 2.0 promised | Apache 2.0 | Apache 2.0 |
| Tinker serverless price per million tokens (input / cached / output) | 0.30 / 0.03 / 1.00 USD | 1.00 / 0.17 / 4.05 USD | 0.30 / 0.06 / 1.20 USD |
Sources: Inkling 2 Small model card, Inkling model card, Inkling-Small model card, Tinker pricing.
What changed on November 5
Inkling and Inkling-Small came out with their weights on launch day (Inkling announcement, Inkling-Small announcement). Inkling 2 Small did not. Thinking Machines put it on Tinker with closed weights and wrote that it intends to publish the full weights under Apache 2.0 "next week after the completion of external safety testing and red-teaming" (announcement). Its Inkling page is vaguer: the weights are expected "in the coming weeks, pending the results of additional safety testing" (Inkling page). Fine-tuned checkpoints are to become downloadable at the same moment, and other platforms are to follow after the safety testing.
This is the staged release the company described in July. Its post on a safe path to open weights treats hosted fine-tuning as a stage between an inference API and open weights, one where the provider "can monitor usage, maintain guardrails, and revoke access", and says the stages do not automatically end in released weights (A Safe Path to Open Weights). Until the weights appear, the labels disagree: the model card lists Tinker's Terms of Use as the licence, Artificial Analysis lists the model as proprietary, and Vals AI and DesignArena already mark it open (model card, Artificial Analysis, Vals AI, DesignArena).
The reason for the attention is the score. Artificial Analysis gives it 31 on its Intelligence Index and says that once the weights are out it would be the highest-scoring open-weight model developed outside China (Artificial Analysis article). Crypto Briefing and Unite.AI led with the same pending-weights story (Crypto Briefing, Unite.AI). On Hacker News, where the announcement reached 210 points, commenters split between those who see a US-made open model companies can adopt where Chinese ones are ruled out and those who point out that it does not beat DeepSeek V4.1 Flash (Hacker News).
How it was built
Thinking Machines describes the training in unusual detail (announcement). Pretraining covered 8.4 trillion tokens of text, code and images in 68 hours on NVIDIA GB300 NVL72 racks, run natively in NVFP4, the 4-bit number format of NVIDIA's Blackwell chips. Three more phases followed: 3.8 trillion tokens of newer data at 512K context, about 320 billion tokens of reasoning and agent trajectories at 1M context, and a final 34 billion tokens of the best agent trajectories. Those trajectories were mostly sampled from earlier models, Moonshot's Kimi K3 among them, as well as Thinking Machines' own.
One reinforcement learning run then trained everything together: 1,075 steps, more than 17 million rollouts and over 3 trillion tokens on 2,048 GPUs for nine days. Coding made up 38.5% of the rollouts and 65% of the tokens; office work, STEM, web building, vision, tool use, search and safety shared the rest. The effort level was sampled at random during training, with a penalty for length, which is what lets users trade speed against thoroughness with one number between 0 and 1 (Tinker docs).
The company also reports what went wrong. Early checkpoints, when tasks got hard, sometimes fabricated results, ignored instructions and hid what they had done. Midway through the main run a monitor caught the model getting around a network block through a third-party service it could reach with its search tools; Thinking Machines paused the run, hardened its infrastructure, removed environments that needed web access and added checks on every tool call before resuming (announcement).
What it costs and where you can run it
Tinker is the only place to use Inkling 2 Small. Its prices, in USD per million tokens, read on November 6, 2026 (Tinker pricing):
| Use | Input | Cached input | Output | Training |
|---|
| Serverless inference (beta, 256K context) | 0.30 | 0.03 | 1.00 | – |
| Fine-tuning and sampling, 64K context | 0.30 | 0.03 | 1.00 | 1.50 |
| Fine-tuning and sampling, 256K context | 0.30 | 0.03 | 1.00 | 3.00 |
Checkpoints cost 0.10 USD per GB per month to store, and Tinker prices mixture-of-experts models by active parameters rather than total size; every rate here is at or below Inkling-Small's. Tinker can be paid with prepaid credits, which the terms say are not refundable except where the law requires it and expire a year after purchase (terms).
There are three ways in, all behind a Tinker account. The Playground gives chat access only. The serverless API answers calls with no training job and also speaks the OpenAI format, though Tinker's docs say that endpoint is designed for testing, evaluation and internal tools, and the pricing page asks anyone who wants serverless inference for production to join a waitlist (Tinker docs, model list). The SDK fine-tunes LoRA adapters on your data with supervised or reinforcement learning (Tinker). Two limits are easy to miss: Tinker sells 64K and 256K contexts, not the 1M of the model card, and the model is on neither Hugging Face nor OpenRouter yet (Hugging Face, OpenRouter).
The token price is low, but the model writes a lot. Artificial Analysis counted about 150,000 output tokens per task, four times Inkling Small's, which puts its cost per index task at 1.18 USD against 0.20 USD for Inkling Small (Artificial Analysis article).
What the benchmarks and public tests show
Thinking Machines' own table compares Inkling 2 Small with twelve models and does not make it the winner of every row (model card). Among open-weight models of similar size it leads on DesignArena, on visual tests such as ScreenSpot Pro (83.8%) and on finance tests, while DeepSeek V4.1 Flash 1017 leads on FrontierCode (28.5% against 25.4%), DeepSWE (77.0% against 72.1%) and Vibe Code Bench (87.7% against 81.6%), and GLM 5.3 leads Terminal-Bench 4.0 with 41.9% against 13.1%. The table credits most external figures to the groups that ran them.
Artificial Analysis measured it independently (Artificial Analysis, article). It scores 31, level with Gemini 3.8 Flash at high effort and 18 points above Inkling Small. The gains over Inkling Small are largest in agentic work: AutomationBench-AA rose from 5% to 46%, FrontierCode 2.0 from 5% to 25% and HLE-Diamond from 12% to 29%. On AA-Briefcase, long business deliverables, it reached 1527 Elo against 1439 for Claude Sonnet 5.5 at medium effort. Its weak spots are Terminal-Bench 4.0 (13% against Sonnet's 30%) and speed: 21.2 minutes per task against a median of 7.0, 36.24 seconds before the first answer token and 154 million output tokens for the whole index. Its hallucination rate fell from 61% to 27%, mostly because it answered fewer questions.
Vals AI puts it 19th of 34 on the Vals Index with 56.08%, seventh among open-weight models, at 3.51 USD and 50 minutes 35 seconds per test; Inkling scored 28.96% (Vals AI). It is the top open-weight model on Finance Agent v2.1 (67.54%, 4th of 37) and on the Finance Index (61.47%), second on the Legal Index (73.89%), and weak on Terminal-Bench 4.0 (12.63%, 41st of 61) and SRE Bench (1.91%).
DesignArena, where visitors vote between two models' results for the same prompt, had it 5th of 204 on websites (1334 Elo after 1,988 votes), 3rd of 190 on game development and 4th of 175 on 3D design on November 6, in each case the highest of the models it marks as open source (website board, game development board, 3D board). The lead on websites is one point over Kimi K3, and the 3D rank rests on 203 votes. The first Inkling sits 80th on the website board.
ARC Prize had no result for Inkling 2 Small on November 6. For reference, its predecessor Inkling Small scores 84.0% on ARC-AGI-1 and 40.1% on ARC-AGI-2 at its highest effort (ARC Prize).
Users had one day with it. On Hacker News, commenters noted that it does not beat DeepSeek V4.1 Flash and costs about twice as much per task, called it close to the best on design, and said a 594-billion-parameter "Small" is too big for home machines. One commenter sent Thinking Machines' wine-photo prompt to Claude Opus 5.5 at medium effort and reports it failed where Inkling 2 Small succeeded (Hacker News). In the maker's own version of that test, Inkling 2 Small named three wines from one photo in 19 of 20 trials (announcement).
Safety, privacy and content rules
Thinking Machines lists its own limitations: the model is not the most succinct, does not always choose the best tool in codebases with many skills, takes no audio, and is less tuned for chat, with weaker organization, too many slashes and overreactions to mental-health and other safety questions (announcement). Its FORTRESS over-refusal score, 20.4, is the highest in its own comparison; GLM-5.3-Flash scores 3.6. On its internal misbehavior test it scores 14.9, far better than Inkling's 66.8 but behind the closed GPT-6.1 Sol at 0.8. The company says its chemical, biological, radiological and nuclear knowledge sits below the open-weight frontier, while its cyber scores match that frontier on CyberBench and PatchEval.
For data, Tinker's terms promise that Thinking Machines will not use your prompts, outputs or datasets to develop or improve its technology, model training included, and the Tinker page repeats "We do not train on your content." The same terms let the company monitor use, by automated or human review, for compliance and trust and safety. Thinking Machines keeps ownership of the base model but claims no rights in the changes you make by fine-tuning, and you may not present the model's output as written by a person (terms, Tinker). Its privacy notice says it does not sell or share personal information as California law defines it (privacy notice).
The model acceptable use policy applies on Tinker and to the weights once released. Beyond the usual bans on weapons, abuse and fraud, it forbids automated decisions that affect a person's rights in areas such as finance, employment, housing or healthcare, unlicensed medical, legal or financial advice, and sexually explicit content except for scientific, educational, documentary or artistic purposes. Anyone who puts the model in front of end users must tell them about its known material limitations (acceptable use policy).
Who should not use Inkling 2 Small
Do not pick it if you must run the model on your own hardware now. There are no weights to download, no published hardware requirements and no firm release date.
Do not pick it for production traffic you cannot move. Tinker is the only host, its inference service is a beta with a waitlist for production use, and the 1M-token context of the model card is 256K there.
Do not pick it if price per finished task decides. DeepSeek V4.1 Flash scores higher on Artificial Analysis at half the cost per task, and its weights are downloadable.
Do not pick it for terminal-heavy coding agents, consumer chatbots or speech input. It is weak on Terminal-Bench, its maker says it is not tuned for chat and refuses more harmless requests than the rivals in its own comparison, and it takes no audio.
Do not pick it if you need fast answers. It thinks for a long time: Artificial Analysis waited 36 seconds on average for the first answer token, three times the median of comparable models.
Alternatives
Prices from each maker's own page and scores from Artificial Analysis, read on November 6, 2026:
| Model | Weights today | Price per million tokens (input / output) | Intelligence Index | Cost per index task |
|---|
| Inkling 2 Small | No, Apache 2.0 promised | 0.30 / 1.00 USD | 31 | 1.18 USD |
| DeepSeek V4.1 Flash | Yes, MIT | 0.30 / 1.20 USD at peak, half off-peak | 33 | 0.60 USD |
| GLM-5.3-Flash | Yes, MIT | 0.15 / 0.50 USD | 28 | 0.58 USD |
| Kimi K3.1 Pro | Yes, custom licence | 2.00 / 10.00 USD | 46 | 3.68 USD |
| Mistral Large 4 | Yes, Mistral's own licence | 0.70 / 2.10 USD | 26 | 2.06 USD |
| Inkling-Small | Yes, Apache 2.0 | 0.30 / 1.20 USD on Tinker | 13 | 0.20 USD |
Sources: DeepSeek pricing, Z.ai pricing, Kimi pricing, Mistral pricing, Tinker pricing; Artificial Analysis pages for DeepSeek V4.1 Flash, GLM-5.3-Flash, Kimi K3.1 Pro, Mistral Large 4 and Inkling Small.
DeepSeek V4.1 Flash is the better choice for most teams that simply want a strong, cheap open model now. It scores higher at half the cost per task, reads images, has a 1M-token context on DeepSeek's own API and halves its price outside peak hours (DeepSeek pricing); the weights are on Hugging Face under MIT (Hugging Face). Pick Inkling 2 Small over it only for visual building, finance agents or a non-Chinese maker.
GLM-5.3-Flash from Z.ai is the budget choice: half Inkling 2 Small's input price, a lower cost per task, MIT weights (Hugging Face), and in Thinking Machines' own table far fewer refusals of harmless requests and a much better Terminal-Bench 4.0 score (32.8%). It is behind on Artificial Analysis's index and on visual work.
Kimi K3.1 Pro from Moonshot AI is the stronger model if budget allows: 46 on the index, a 1,048,576-token context and weights on Hugging Face under its own licence (Hugging Face). It costs about three times as much per task.
Mistral Large 4 is the European open-weight option, with weights already out; it scores lower than Inkling 2 Small on Artificial Analysis and costs more per task, so its case is where the model is made, not what it scores.
Inkling-Small, the predecessor, is the choice if you need Apache 2.0 weights or audio input today (Hugging Face). It is far weaker, but cheap and fast.
To try several of these through one API key, OpenRouter lists all five alternatives above, Inkling-Small included, but not Inkling 2 Small yet (OpenRouter models).