Side-by-side sourced facts
Both columns quote the same reviewed Epoch AI catalog artifact. The LMArena row quotes the separately pinned snapshot published July 12, 2026 and appears only for an exact source model match. Missing values stay missing.
| Field |  Grok 4.1 |  Pixtral Large |
|---|---|---|
| Developer | xAI | Mistral |
| Released | November 17, 2025 | November 18, 2024 |
| Domains | Language | Multimodal, Language, Vision |
| Tasks | Language modeling/generation, Question answering, Mathematical reasoning, Quantitative reasoning, Code generation | Vision-language generation, Visual question answering, Mathematical reasoning, Character recognition (OCR), Language modeling/generation, Question answering |
| Access | API access | Open weights (non-commercial) |
| Weights | Closed weights | Open weights |
| LMArena rating | No exact model match in this snapshot | No exact model match in this snapshot |
| Context window | Not reported by this source | Not reported by this source |
| Price | Not reported by this source | Not reported by this source |
Common questions
Is Grok 4.1 better than Pixtral Large?
No published ranking covers both models in Model Gauntlet's approved sources: the LMArena snapshot published July 12, 2026 has no exact match for either model. The Epoch catalog documents what each model is for; it does not measure which is better.
Which is newer, Grok 4.1 or Pixtral Large?
According to Epoch AI's model catalog (retrieved July 14, 2026), Grok 4.1 is newer: it was released November 17, 2025, while Pixtral Large was released November 18, 2024.
Can you self-host Grok 4.1 or Pixtral Large?
Only Pixtral Large, according to the Epoch records: it ships open weights limited to non-commercial use, while Grok 4.1 is a closed-weight model with API access only.
Evidence boundary
This page declares no winner beyond what a named source's numbers say. Where the sources do not overlap, the page says so plainly. Task quality, price, context window, latency, and reliability are not reported by these sources, so no claim about them appears here. Both models still require hands-on evaluation on your own workload.