Inkling 2 Small vs Mistral Large 4
Both models appear in the same evaluation table on the official model card, which is the only place they are measured under identical conditions and reported by the same party.
Across 46 benchmarks where both models report a score, Inkling 2 Small is higher on 39 and Mistral Large 4 is higher on 6 , with 1 exact ties. Inkling 2 Small has 594B total parameters with 17B active, and is priced at $1 per million output tokens on Tinker. Its largest margin is on DesignArena Game Dev. Its largest deficit is on AutomationBench v1.0.6.
Head to head
- Position
- Inkling 2 Small: the subject of this site · Mistral Large 4: open weights, larger class
- Bechmarks scored
- 46 shared rows, out of 71 total
- Record
- 39 wins · 6 losses · 1 ties
Largest differences
| Benchmark | Inkling 2 Small | Mistral Large 4 | Gap |
|---|---|---|---|
| DesignArena Game Dev · Visual Creation | 1,428 | 1,225 | + 203 |
| AA-Briefcase v1.1 · Knowledge Work | 1,527 | 1,348 | + 179 |
| DesignArena 3D Design · Visual Creation | 1,401 | 1,249 | + 152 |
| Code Arena · Visual Creation | 1,653 | 1,544 | + 109 |
| DesignArena Website · Visual Creation | 1,340 | 1,256 | + 84 |
| ScreenSpot Pro · Visual Perception | 83.8% | 42.3% | + 41.5 |
| BabyVision · Visual Perception | 68.6% | 42.3% | + 26.3 |
| PMPP-Hard · Autoresearch | 494.27 | 478.39 | + 15.88 |
| Chartography · Visual Reasoning | 36.7% | 21.2% | + 15.5 |
| JobBench · Knowledge Work | 70.5% | 55.2% | + 15.3 |
Every shared benchmark
| Benchmark | Inkling 2 Small | Mistral Large 4 | Leader |
|---|---|---|---|
| Third Party Index | |||
| AA Index (v5.0) | 30.6 | 26.1 | Inkling 2 Small |
| Vals Index (v2.2) | 56.1% | 56.5% | Mistral Large 4 |
| Coding | |||
| FrontierCode v2.0 | 25.4% | 15% | Inkling 2 Small |
| DeepSWE v1.1 | 72.1% | 61.7% | Inkling 2 Small |
| Terminal Bench Science v0.1 | 5.7% | 4.3% | Inkling 2 Small |
| Terminal Bench v4.0 | 13.1% | 24.8% | Mistral Large 4 |
| Vibe Code Bench v1.1 | 81.6% | 81.6% | tie |
| Sudo L7 | 14.1% | 13.9% | Inkling 2 Small |
| SWEBench Multilingual | 81.2% | 79.7% | Inkling 2 Small |
| SWEBench Verified | 82.2% | 80% | Inkling 2 Small |
| Autoresearch | |||
| PMPP-Hard · AUC (task-hours, 0–10h) | 494.27 | 478.39 | Inkling 2 Small |
| Search and Tools | |||
| BrowseComp · with context management | 90.2% | 84.5% | Inkling 2 Small |
| MCP Atlas · Full / Public | 85.4% | 84.5% | Inkling 2 Small |
| Knowledge Work | |||
| AA-Briefcase v1.1 | 1,527 | 1,348 | Inkling 2 Small |
| JobBench | 70.5% | 55.2% | Inkling 2 Small |
| Enterprise Operations | |||
| AutomationBench v1.0.6 | 45.6% | 57.8% | Mistral Large 4 |
| Finance Work | |||
| Finance Agent v2.1 | 67.5% | 61.7% | Inkling 2 Small |
| Excel Modeling Benchmark v1.1 | 69.5% | 68.7% | Inkling 2 Small |
| Tax Agent Bench | 65.8% | 67.5% | Mistral Large 4 |
| DAYJOB (Finance) | 14.3% | 4.8% | Inkling 2 Small |
| Apex-Accounting v1.1 · pass@1 / mean score | 14.5% | 12% | Inkling 2 Small |
| Legal Work | |||
| Harvey LAB v1.1 | 21.7% | 15.8% | Inkling 2 Small |
| Legal Research Bench | 46.6% | 35.6% | Inkling 2 Small |
| Visual Creation | |||
| DesignArena Website | 1,340 | 1,256 | Inkling 2 Small |
| DesignArena Game Dev | 1,428 | 1,225 | Inkling 2 Small |
| DesignArena 3D Design | 1,401 | 1,249 | Inkling 2 Small |
| Code Arena | 1,653 | 1,544 | Inkling 2 Small |
| Visual Perception | |||
| ScreenSpot Pro · original / with Python | 83.8% | 42.3% | Inkling 2 Small |
| BabyVision · original / with Python | 68.6% | 42.3% | Inkling 2 Small |
| Visual Reasoning | |||
| GDP.pdf | 25.2% | 17.6% | Inkling 2 Small |
| MMMU Pro (Standard 10) | 81.6% | 78% | Inkling 2 Small |
| Charxiv RQ · original / with Python | 85.6% | 83.2% | Inkling 2 Small |
| Chartography · original / with Python | 36.7% | 21.2% | Inkling 2 Small |
| MathVision | 92.8% | 88.9% | Inkling 2 Small |
| STEM Reasoning | |||
| CritPt v2 · Full / Hard | 55.7% | 50.3% | Inkling 2 Small |
| HLE-Diamond | 28.7% | 17.3% | Inkling 2 Small |
| GPQA Diamond | 93% | 89.7% | Inkling 2 Small |
| Long Context | |||
| LongBench v2 | 70.6% | 65.7% | Inkling 2 Small |
| Factuality | |||
| AA-Omniscience v1.1 · score | 16.1 | 5.1 | Inkling 2 Small |
| AA-Omniscience v1.1 · non-hallucination rate | 73.3% | 71.8% | Inkling 2 Small |
| Instruction Following | |||
| ComplexConstraints v1.0 | 44.5% | 36.5% | Inkling 2 Small |
| Safety | |||
| DistressBench | 82% | 77% | Inkling 2 Small |
| FORTRESS (adversarial) | 83.8% | 76% | Inkling 2 Small |
| FORTRESS (benign) | 79.7% | 91.8% | Mistral Large 4 |
| StrongREJECT (none) | 97.1% | 99% | Mistral Large 4 |
| Cybersecurity | |||
| PatchEval | 79.6% | 76.1% | Inkling 2 Small |
Reading the comparison honestly
These numbers come from one table published by Thinking Machines, using third-party evaluations where those exist. That makes them comparable to each other, but it does not make them neutral: the vendor chose which benchmarks to run and how to present them, and a benchmark that a model loses has a way of not appearing.
The models are also not the same shape of thing. Mistral Large 4 sits in the open weights, larger class tier of the card's own grouping, while Inkling 2 Small is the subject of this site. A win against a smaller model is less meaningful than a win against a larger one, and the table does not adjust for that.