Inkling 2 Small is Thinking Machines's open-weight AI model, released 05/11/2026. Inkling 2 Small is a 594B-parameter mixture-of-experts model with 17B active, accepting image input and supporting variable effort. Thinking Machines said the full weights would be released under Apache 2.0 after external safety testing and red-teaming (the page notes weights on Hugging Face on 12 November 2026). Its comparison table covers GLM-5.3-Flash, MiMo V2.6, MiniMax-M3.1-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open, GLM 5.3, Mistral Large 4, Kimi K3, GPT-6 Luna, Gemini 3.8 Flash and Claude Haiku 5.5. The page prints some figures inconsistently (Intelligence Index 31 against 30.6, DesignArena 1331 against 1340). Our article on this model is the sibling Inkling 2 Mini piece.
Visit Inkling 2 SmallAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does Inkling 2 Small score on independent benchmarks?
Inkling 2 Small ranks #20 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 30.6, as measured on 22 November 2026.
Where does Inkling 2 Small rank against other models?
How does Inkling 2 Small do on individual tests?
| Evaluation (run by Artificial Analysis) | Inkling 2 Small | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 13.1% | 67.2% |
| FrontierCode 2.0 | 25.4% | 48.8% |
| Humanity's Last Exam | 44.1% | 74.3% |
| GDPval-AA (Elo) | 1630 | 1866 |
How has Inkling 2 Small's score changed over time?
We have tracked Inkling 2 Small on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the Inkling 2 Small system card report?
Inkling 2 Small is a 594B-parameter mixture-of-experts model with 17B active, accepting image input and supporting variable effort. Thinking Machines said the full weights would be released under Apache 2.0 after external safety testing and red-teaming (the page notes weights on Hugging Face on 12 November 2026). Its comparison table covers GLM-5.3-Flash, MiMo V2.6, MiniMax-M3.1-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open, GLM 5.3, Mistral Large 4, Kimi K3, GPT-6 Luna, Gemini 3.8 Flash and Claude Haiku 5.5. The page prints some figures inconsistently (Intelligence Index 31 against 30.6, DesignArena 1331 against 1340). Our article on this model is the sibling Inkling 2 Mini piece.
These are Thinking Machines's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Source: Inkling 2 Small (05/11/2026). Last checked 22/11/2026.
Coding
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| FrontierCode v2.0Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 25.4% | n/a | n/a | n/a |
| DeepSWE v1.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 72.1% | 69% | 68.5% | 66.6% |
| SWE-Bench Pro v2 (Hard)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 86.3% | 84.3% | 88.2% | 88.7% |
| SWE-Bench Pro v2 (Full)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 97.7% | n/a | n/a | n/a |
| SWE-Atlas v1.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 61.9% | n/a | n/a | n/a |
| Vibe Code Bench v1.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 81.6% | n/a | n/a | n/a |
| Terminal Bench Science v0.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 5.7% | 9.5% | 8.1% | 8.6% |
| PostTrainBench v1.2Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 31.9% | n/a | n/a | n/a |
| PMPP-Hard (AUC, 0 to 10h)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 494.27 AUC | n/a | n/a | n/a |
Agents & tool use
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| AA-Briefcase v1.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 1527 Elo | n/a | n/a | n/a |
| GDPval-AA v2.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 1630 Elo | 1657 | 1543 | 1444 |
| JobBenchEffort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 70.5% | n/a | n/a | n/a |
| EnterpriseOps-GymEffort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 46.8% | n/a | n/a | n/a |
| Finance Agent v2.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 67.5% | n/a | n/a | n/a |
| Apex-Accounting v1.1 (pass@1)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 14.5% | n/a | n/a | n/a |
| BrowseComp (with compaction)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 90.2% | 75.1% | 88.4% | 83.6% |
| DeepSearchQA (with compaction)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 90.9% | n/a | n/a | n/a |
Knowledge & science
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| HLE-DiamondEffort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 28.7% | 17.9% | 26.9% | 22.2% |
| CritPt v2 (Full)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 55.7% | n/a | n/a | n/a |
| CritPt v2 (Hard)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 35% | n/a | n/a | n/a |
| AA-Omniscience scoreEffort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 16.1 points | n/a | n/a | n/a |
| AA-Omniscience non-hallucination rateEffort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 73.3% | n/a | n/a | n/a |
Reasoning & maths
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| ARC-AGI-1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 96.5% | n/a | n/a | n/a |
| ARC-AGI-2Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 81.7% | n/a | 60.4% | 59.3% |
Long context
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| AA-LCR v1.1Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 83.3% | 79.7% | 88.7% | 83.3% |
| LongBench v2Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 70.6% | n/a | n/a | n/a |
Multimodal
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| MMMU Pro (Standard 10)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 81.6% | n/a | 80.5% | 79.7% |
| PerceptionBenchEffort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 64.7% | n/a | n/a | n/a |
| ScreenSpot Pro (original)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 83.8% | n/a | 73.8% | 81.1% |
| ScreenSpot Pro (with Python)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 92% | n/a | 86.1% | 87.7% |
| CharXiv RQ (original)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 85.6% | n/a | n/a | n/a |
| CharXiv RQ (with Python)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 89.1% | n/a | n/a | n/a |
| BabyVision (original)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 68.6% | n/a | n/a | n/a |
| BabyVision (with Python)Effort 0.99 (xhigh), temperature 1.0, 1M-token maximum trajectory | 93.6% | n/a | n/a | n/a |
Other
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| DesignArena WebsiteTable value; the chart on the same page shows 1340 | 1331 Elo | n/a | n/a | n/a |
| Artificial Analysis Intelligence IndexBenchmark table value; the page intro rounds to 31 | 30.6 points | n/a | n/a | n/a |
Safety & honesty
| Benchmark | Inkling 2 Small | GLM 5.3 | Kimi K3 | GPT-6 Luna |
|---|---|---|---|---|
| StrongREJECTHigher is better | 97.1 points | n/a | n/a | n/a |
| FORTRESS average risk scoreLower is better | 16.2 points | n/a | n/a | n/a |
| FORTRESS overrefusal scoreLower is better | 20.4 points | n/a | n/a | n/a |
What did safety testing find for Inkling 2 Small?
Safety level: Thinking Machines' own evaluations; no external framework level stated
- DistressBench 82.0 (higher is better) and PatchEval 79.6 (higher is better); CyberBench v1.1 72.7.
- Internal misbehaviour rate 14.9, covert misbehaviour 11.8 and destructive actions 4.7 (all lower is better).
- Apollo Research covert-action-with-goal result is 0.
- The page states Inkling 2 Small trails GPT-6.1 Sol on its alignment evaluations.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our Inkling 2 Small deep dive.
Thinking Machines Model Timeline
Not published as a product limit; evaluations used a 1M-token maximum trajectory (256K for RL training) context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
