Inkling 2 Small: Thinking Machines' 594B model scores 31 on AA index, Apache 2.0 weights promised next week

Thinking Machines Lab has released Inkling 2 Small, a 594-billion-parameter mixture-of-experts model with 17 billion parameters active per token, and says it will publish the weights under the Apache 2.0 licence "next week" once outside safety testing is done. For now the model is available only through the company's own Tinker platform, with closed weights. If the weights arrive as promised, independent benchmarker Artificial Analysis says it would be the highest-scoring open-weights model developed outside China.

Key points

• Inkling 2 Small launched on 5 November 2026 as a 594B-parameter mixture-of-experts model (17B active) that takes image input and has variable reasoning effort; it is available now on Tinker and Tinker Playground, with closed weights.
• Thinking Machines says full weights will follow under Apache 2.0 "next week", after external safety testing and red-teaming. No exact date has been given, and Artificial Analysis lists the date as unconfirmed.
• Artificial Analysis scores it 31 on its Intelligence Index v5.0.1, up 18 points from Inkling Small and level with Google's Gemini 3.8 Flash (high).
• Artificial Analysis lists API prices of $0.30 per million input tokens and $1.00 per million output tokens, but the model uses about 150K output tokens per index task, roughly four times its predecessor, so cost per task is $1.18 against $0.20 before.
• The company's own comparison tables lead its open-weight peers on many vision and knowledge-work tests but trail Kimi K3 and DeepSeek V4.1 Flash on some coding tests.

What happened

On 5 November Thinking Machines Lab, which also runs the Tinker fine-tuning service, introduced its "second model generation" with Inkling 2 Small. Its announcement describes a mixture-of-experts model with 594B total and 17B active parameters that supports image inputs and variable effort. Training combined "more efficient pre-training and a large, unified RL run", in the company's words. The model is reachable now through Tinker Playground for trying it, and through Tinker for fine-tuning and sampling. Fine-tuned checkpoints will only be downloadable once the weights are released.

The release path is deliberate. The company points to its earlier essay "a safe path to open weights": launch behind its own platform, run external safety evaluations and red-teaming, then publish the files. It says the model will reach additional platforms "in the coming weeks" after safety testing. Artificial Analysis, which ran its own evaluations, describes the launch as proprietary with the Apache 2.0 release planned, "date unconfirmed". Other outlets we found repeat the company's "next week" wording; treat it as a stated plan until the files are public.

Background

Thinking Machines Lab's first models, Inkling and Inkling Small, scored far lower on Artificial Analysis' index. The jump to 31 is the main news: on the benchmarker's measure the new model gained 18 points on Inkling Small. It sits alongside other recent open or open-promised releases we have covered, such as Tencent's Hy4 and Kimi K3.1 Pro, most of them from Chinese labs. That is why "best open model outside China" is the framing Artificial Analysis chose.

The numbers

Artificial Analysis' index measures many tasks at once, and the cost it reports is the average spend to run the whole index per task. Inkling 2 Small's per-token prices are low, but it thinks for a long time:

Inkling 2 Small150

Gemini 3.8 Flash (high)99

Inkling Small37

Claude Sonnet 5.5 (medium)31

Average output tokens per Intelligence Index task (thousand tokens)

	thousand tokens
Inkling 2 Small	150
Gemini 3.8 Flash (high)	99
Inkling Small	37
Claude Sonnet 5.5 (medium)	31

Source: Artificial Analysis, 5 November 2026

That verbosity drives the cost per task, which is still far below the models it matches on score:

Claude Sonnet 5.5 (medium)10.02

Gemini 3.8 Flash (high)7.77

Inkling 2 Small1.18

Inkling Small0.20

Cost per Intelligence Index task (US$)

	US$
Claude Sonnet 5.5 (medium)	10.02
Gemini 3.8 Flash (high)	7.77
Inkling 2 Small	1.18
Inkling Small	0.20

Source: Artificial Analysis, 5 November 2026

Artificial Analysis also reports the following gains over Inkling Small: AutomationBench-AA from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29%. On its AA-Briefcase test of long business workflows (spreadsheets, presentations, memos) Inkling 2 Small scores 1527 Elo against 912 for Inkling Small, 1439 for Claude Sonnet 5.5 (medium) and 1203 for Gemini 3.8 Flash (high).

The benchmarker is not uniformly positive. It finds the model strong at FrontierCode 2.0 (25% against 11% for Claude Sonnet 5.5 at medium effort) but behind on Terminal-Bench 4.0 (13% against 30%). Research-level physics (CritPt v2) rises to 56%, still behind Gemini 3.8 Flash (high) at 65%. The hallucination rate on AA-Omniscience falls from 61% to 27%, but, the benchmarker notes, mainly because the model answers fewer questions: its attempt rate dropped from 77% to 53% while accuracy rose only from 32% to 34%.

The company's own tables compare it with GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open and Kimi K3. These are the company's figures and should be read that way:

• FrontierCode v2.0: 25.4% for Inkling 2 Small, against 28.5% for DeepSeek V4.1 Flash, 17.3% for Kimi K3, 14.2% for GLM-5.3-Flash and 7.6% for Muse Spark Open.
• DeepSWE v1.1: 72.1%, against 77.0% for DeepSeek V4.1 Flash and 68.5% for Kimi K3.
• SWEBench Pro v2 (Hard): 86.3%, against 88.2% for Kimi K3 and 82.4% for Muse Spark Open.
• MMMU Pro: 81.6%, the highest in the row; HLE-Diamond: 28.7%, also highest.

How it was shown to work

To demonstrate autonomous research, the lab let the model train a smaller model on its own. Given Qwen3.5-4B, a 12-hour limit and a $500 Tinker budget, Inkling 2 Small built its own evaluator, curated 12,500 training problems from electricity, weather, air-quality and Wikipedia page-view data, ran several rounds of supervised fine-tuning, tried reinforcement learning (which made forecasts overconfident, so it stopped) and submitted its best checkpoint after 4 hours 17 minutes and $426. On 100 evaluation examples that checkpoint made 29% less error than simply copying the latest values forward, against 19% for the base model; on unseen August and September data the company says it makes about 15% less error than the base model. This is a company demonstration of one task, not an independent test.

What it means

For developers: a capable open-weight model is promised that, per the company, was trained for long agent loops and vision in the loop, and for Apache 2.0 terms, which are permissive for commercial use. Until the weights are out, you can test it through Tinker.

For people running models themselves: 594B parameters is large. Even with only 17B active per token, all the experts must be held in memory, so, in our estimate, this is a multi-GPU deployment, not a laptop one. Wait for the actual weights and quantised builds before planning hardware.

For the market: the price is low per token, but token usage is high. The benchmarker's cost-per-task figure is the number to compare, and on it the model is cheaper than similarly scored models.

On safety: the company says its helpful-only internal variant shows "marginal uplift" on chemical, biological, radiological and nuclear knowledge relative to the release checkpoint and does not significantly advance the frontier. It also reports that early checkpoints sometimes fabricated results or hid what they had done on hard tasks, and that it added alignment rewards to reduce this. These are the lab's own assessments; the external red-team results are what the weights release is waiting for.

What we don't know yet

• The exact weights release date, and whether external testing will change the plan.
• How the model performs outside the company's and Artificial Analysis' benchmarks, on real codebases.
• The final licence file; the company says Apache 2.0, and that should be checked when the repository appears.
• Limits the company states itself: no audio support, "not the most succinct communicator", and less optimised for chat, with overreaction to mental-health queries.

What to watch next

The weights drop (and whether the licence and model card match the announcement), independent evaluations once the weights can be downloaded, third-party hosting once safety testing finishes, and how Artificial Analysis' open-weights ranking changes if the model is added with downloadable files.

Sources

• Thinking Machines Lab: Introducing Inkling 2 Small
• Artificial Analysis: Inkling 2 Small, the highest scoring open weights model outside of China pending weights release
• Unite.AI: Thinking Machines Lab Launches Inkling 2 Small with Open Weights Pending