Thinking Machines' Inkling 2 Small claims open-weight crown at six times the cost

Thinking Machines has released Inkling 2 Small, a 594-billion-parameter model that scores 31 on the Artificial Analysis Intelligence Index — an 18-point jump over its predecessor and roughly on par with Google's Gemini 3.8 Flash. The startup, founded by former OpenAI CTO Mira Murati, claims the model is now the strongest Western open-weight model. The claim rests on vendor-supplied benchmark numbers, and the performance gains carry a clear price: each task costs nearly six times as much as before.

The company positions Inkling models as base models for fine-tuning, meaning developers adapt them for specialized applications rather than using them as general-purpose chatbots. Inkling 2 Small is the second model generation from the startup, which has made open weights a central part of its pitch.

What are the architecture and headline numbers?

Inkling 2 Small is a mixture-of-experts model (a design where only a fraction of the network activates per query) with 594 billion total parameters, of which 17 billion are active at any given time. It accepts text and images as input. Users can adjust its thinking depth, trading speed for thoroughness on a per-task basis.

On the Artificial Analysis Intelligence Index, the model scores 31, according to Thinking Machines. The predecessor, Inkling Small, scored 13. That places the new model roughly level with Google's Gemini 3.8 Flash.

The largest gains appear in agentic tasks, where the model independently executes multi-step workflows. The numbers, as reported by the company:

• AutomationBench-AA: 5 percent to 46 percent
• FrontierCode 2.0: 5 percent to 25 percent
• AA-Briefcase (knowledge work): 1,527 Elo, up 615 points

The claimed hallucination rate falls from 61 percent to 27 percent. That improvement comes with an asterisk: the model answers fewer questions overall. Its response rate drops from 77 percent to 53 percent. A model that declines to answer nearly half the time will hallucinate less by construction.

How much does it cost to run?

The improved output comes partly from generating far more text. Inkling 2 Small produces roughly 150,000 output tokens per task on average — about four times as many as its predecessor — according to Artificial Analysis.

At $1.18 per task, it costs nearly six times as much as the previous generation. That figure still undercuts Gemini 3.8 Flash ($7.77 per task) and Claude Sonnet 5.5 ($10.02) by a wide margin. Artificial Analysis notes it lands closer to the price range of low-cost Chinese models. The cost comparison matters because benchmark parity with frontier models means little if inference economics shift the equation; here, the advantage remains with the newcomer.

What do the coding results show?

Thinking Machines says it trained the model with reinforcement learning (a technique that rewards desired outputs through trial and error) on hundreds of thousands of long-horizon coding tasks. An automated evaluation system checked whether the generated code worked and how the model approached each task.

On the coding benchmarks the company reported:

• FrontierCode 2.0: 25.4 percent
• DeepSWE: 72.1 percent
• SWEBench Pro v2 (Hard): 86.3 percent

The new DeepSeek-V4.1-Flash still leads on most of those tests, with 28.5 percent on FrontierCode and 77 percent on DeepSWE. The Chinese model's edge is narrow but consistent, which tempers the claim of overall open-weight leadership.

Where does the model lead on vision?

Thinking Machines says the model performs best in text recognition, infographic interpretation, and visual reasoning. On GDP.pdf, a document benchmark testing whether models can answer questions from complex PDF documents, it scores 25.2 percent — the highest among the compared models. On MMMU Pro, a benchmark for multimodal understanding of academic tasks involving diagrams and figures, it reaches 81.6 percent.

On PerceptionBench, which tests visual perception in multimodal models, it leads with 64.7 percent. For visual creation, the model earns a 1,340 Elo rating on DesignArena Website, where users rate AI-generated websites in head-to-head comparisons. That is the highest score among open-weight models. The company also reports progress in 3D modeling, without publishing a supporting benchmark.

In an unusual demonstration, the team pitted the model against its own researchers in a blind wine tasting. Working from a photo alone, Inkling 2 Small correctly identified three wines in 19 out of 20 attempts. The stunt is memorable; it is also unaudited and self-reported.

How compact was the training run?

The training run is notably small relative to competitors. Across all phases, the model processed about 12 trillion tokens, according to Thinking Machines — far fewer than comparable rivals, the company says, without naming specific comparison figures. The reinforcement-learning phase after base training required only 3 trillion tokens on 2,048 GPUs over nine days.

If those numbers hold up under independent scrutiny, they would suggest meaningful efficiency gains in post-training. Nothing in the announcement has yet been verified outside the company.

What are the model's limits?

Thinking Machines discloses a candid list of weaknesses. The model sometimes communicates inefficiently. It does not always pick the best tools for large codebases — which the company flags as likely its bigger issue. It does not support audio, and it is less optimized for chat than consumer-facing models. It can also overreact to health-related topics and produce cluttered output.

For now, the model is available only as a proprietary offering through the company's Tinker platform. The open-weight claim therefore describes a future state, not the current release.

When do the weights actually open up?

Thinking Machines plans to release the weights under an Apache 2.0 license once external safety testing wraps up. Apache 2.0 is one of the most permissive open-source licenses, and it would give Inkling an advantage over more restrictive open-weight licenses such as those used by Meta or Mistral.

The company says Inkling 2 Small belongs to a family of models in different sizes, with more variants planned for release soon. Until the weights ship and third parties reproduce the benchmark results, the strongest-Western-open-weight claim remains a promise backed by the company's own numbers.