Inkling 2 Small shakes up the openness of Western AI models
Thinking Machines Lab releases Inkling 2 Small, an open model aiming for the best scores outside China. The system is still to receive its weights next week after a security testing phase.
Thinking Machines Lab, a start-up founded by former OpenAI CTO Mira Murati, has unveiled Inkling 2 Small, an open AI model aiming to compete with Chinese benchmarks in the open weight model field. According to Artificial Analysis, it could become the best open model developed outside China once its weights are released next week. This signal matters for the market, as the open source battle in AI is no longer just about raw performance but also about speed of deployment and practical use cases.
A model designed for coding, designing, and staying usable
Inkling 2 Small features 594 billion parameters, accepts image inputs but produces only text. Thinking Machines says it evaluated its capabilities on coding tasks, as well as on uses related to web design and game design. The company’s stated result: “a model that is both stronger at coding and easier to use inside real agent loops”. In other words, the model is designed to fit into AI agents that chain actions without drifting too much.
On the Artificial Analysis Intelligence Index, which aggregates tests on coding, science, and general knowledge, Inkling 2 Small scores 31 points. This places it above the average in its category and tied with Gemini 3.8 Flash, according to the source. However, the ranking leaders remain higher, with Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59.
Why releasing the weights changes the game
The company plans to release the model weights next week. This practice has become common: publishers first keep control for a few days or weeks to conduct cybersecurity and product safety tests with external reviewers before releasing the model. The word “open” therefore has several layers of meaning, and the industry loves its nuances, even when it claims to move fast.
This schedule places Thinking Machines within a broader movement. Mistral recently followed a similar sequence with Le Chonk, its open code-oriented model. According to the source text, this French model now matches the pace of OpenAI GPT-6.1 Terra on code, which remains rare for a Western open model. In this context, Inkling 2 Small is not just another launch: it is part of an attempt to regain initiative against Chinese models that often dominate open weight rankings.
Clear progress, but not without trade-offs
Artificial Analysis considers Inkling 2 Small solid, while pointing out two flaws: it is slow per task and very verbose. Its price is also described as “somewhat expensive,” at 0.30 dollars per million input tokens, while the median in its category is 0.24 dollars. In other words, the model improves, but the cost remains an issue.
The gains compared to the previous version are nevertheless significant. According to Artificial Analysis, automation capabilities increase by 41% and coding capabilities by 20%. On the AA-Omniscience v1.1 test, the hallucination rate drops from 61% to 27%, while accuracy rises from 32% to 34% and the attempt rate falls from 77% to 53%. These figures suggest a more reliable model, even if competition remains fragmented: few models dominate all benchmarks simultaneously, and this is where product teams start counting their tokens [to be verified].
Key points
- Inkling 2 Small targets open weight leaders outside China.
- 594 billion parameters, with image input and text output.
- 31 points on the Artificial Analysis Intelligence Index.
- Weights announced for next week.
- Hallucinations: 61% to 27% on AA-Omniscience v1.1.
- Thinking Machines Lab is linked to Mira Murati.
By the numbers
- 594 billion — model parameters, according to Thinking Machines Lab.
- 31 points — Inkling 2 Small’s score on the Artificial Analysis Intelligence Index.
- 0.30 dollar — cost per million input tokens announced by the source.
- 41% — gain in automation capabilities compared to Inkling Small.
- 27% — hallucination rate on AA-Omniscience v1.1, down from 61% before.