Thinking Machines Claims a Benchmark Crown It Won't Let You Check
Artificial Analysis says Inkling 2 Small matches Gemini 3.8 Flash and would be the best open model outside China. The weights are still locked behind Mira Murati's own platform.

Mira Murati's Thinking Machines Lab announced Thursday, Nov. 5 that it had built the best open-weights model made outside China. Then it told you that you cannot have it.
The benchmark number is real. The "open" part is a plan with no enforceable date.
Thinking Machines released Inkling 2 Small, scoring 31 on the Artificial Analysis Intelligence Index. That would make it the highest-scoring open-weights model developed outside China once the weights ship. It has 594 billion parameters with 17 billion active per token, reads text and images, and matches Gemini 3.8 Flash on Artificial Analysis's leaderboard. Today it runs only on Tinker, the lab's own platform, in closed form.
Artificial Analysis is a legitimate third party here, not a Thinking Machines subsidiary grading its own work. The 31 represents a real jump: up from 13 for Inkling Small.
Translation of the official line: "Inkling 2 Small will initially be available through Tinker with closed weights, following our outlined safe path to open weights. We intend to make full weights available under Apache 2.0 after external safety testing and red-teaming." Plain English: trust the score, not the file. The file ships when we decide.
The Crown Has an Asterisk
Among open models, only Kimi K3.1 Pro (46), DeepSeek V4.1 Flash, and GLM-5.3 (both 32) score higher, and all three are Chinese. Mistral Large 4 published its weights on Oct. 29 and held the non-China crown at 26 points for a week. Inkling 2 Small clears that by five points using far less compute per token: 17 billion active parameters versus 49 billion for Mistral.
Artificial Analysis flagged one wrinkle. The score jumped 18 points while the model generates roughly four times as many output tokens per task. Most improvement comes from the model attempting fewer questions, its attempt rate falling from 77% to 53%. A model that skips more questions and writes longer answers to the rest isn't automatically smarter. It might just be more selective.

What Verification Actually Requires
You cannot run Inkling 2 Small on your own hardware. Every number in this story, including Artificial Analysis's own run, was generated against a hosted endpoint the lab controls, not a file on an outside researcher's GPUs where no one can throttle it or swap it between runs. The lab's own benchmark claims—first place on AA-AnalystAgent and Dayjob:Finance, second on Vals Finance Index and DiligenceBench—are lab-run with no independent verification yet.
The promise comes pre-hedged: full weights under Apache 2.0 next week, pending completed external safety testing and red-teaming. Pending. Not scheduled. Not guaranteed.
None of this means Thinking Machines is lying. It means the company controls every variable that would let you check.
A startup run by a former OpenAI CTO, mid funding cycle, has every reason to let the score travel for a news cycle before the file travels at all.
What to Watch
Whether the Apache 2.0 weights actually land the week of Nov. 9 as promised. If a downloaded checkpoint matches the Tinker score, the crown is real. If the date slips or the file underperforms the hosted demo, the story becomes the gap itself. Watch whether Artificial Analysis flips the model's leaderboard tag from proprietary to open once files exist. That flag, not the press release, is the only honest scoreboard here.
- Introducing Inkling 2 Small - Thinking Machines Lab
- Inkling 2 Small - Intelligence, Performance & Price Analysis
- Thinking Machines Releases Inkling 2 Small, To Become Top Scoring Non-Chinese Model
- Murati's Inkling 2 Small Scores 31; Output Price Down 75%
- Inkling 2 Small scores 31 and takes Europe's open-weight crown, if the weights arrive next week
- Inkling 2 Small Is Here — the Weights Aren't. What to Do Now
- Inkling 2 Small: 594B Model, Apache Weights Next Week
- Thinking Machines Releases Inkling 2 Small, a 594B Agent-First Model
- Inkling Small lands within a point of Inkling on the Artificial Analysis Intelligence Index with less than a third of the parameters
- Inkling 2 Small: Thinking Machines' new model improves on Inkling 1
- Inkling 2 Small
- Thinking Machines releases Inkling 2 Small, scores 31 on AI Index
- Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active - MarkTechPost
- Inkling 2 Small Benchmarks, Cost and Capabilities
- Thinking Machines launches Inkling 2 Small, open weights due next week
- Inkling: Our Open-Weights Model - Thinking Machines Lab
- Artificial Analysis on X: "Thinking Machines has released Inkling 2 Small, scoring 31 on the Artificial Analysis Intelligence Index - their plan to release its weights would make it the highest-scoring open weights model developed outside China based on current rankings Inkling 2 Small is launchin… / X