MiMo V2.6 Pro vs Mistral Large 4
Mistral Large 4 and MiMo V2.6 Pro score 56.8% vs 50.3% averaged across five Vision Evals tasks (#51 and #64 of 73 models), so Mistral Large 4 leads overall.
- Mistral Large 4 wins 4 of the 5 tasks, with its widest lead on object detection (59.3% vs 42.0%).
- MiMo V2.6 Pro is about a third of the cost: $0.0011 vs $0.0033 overall average cost per sample.
- MiMo V2.6 Pro is faster: 9.5s vs 31.2s per sample.
Roboflow Vision Evals. Full breakdown below.
Compare MiMo V2.6 Pro vs Mistral Large 4 live
Run the same image across every model that supports a task and compare their outputs side-by-side.
Detect and compare bounding boxes across models on the same image.
Upload an image
Drag and drop an image here, or click to browse
MiMo V2.6 Pro vs Mistral Large 4 on Vision Evals
Roboflow Vision Evals. Full breakdown below.
MiMo V2.6 Pro vs Mistral Large 4 Comparison Table
Evals updated November 19, 2026Pricing updated November 22, 2026
| Property | MiMo V2.6 Pro | Mistral Large 4 |
|---|---|---|
| Organization | Xiaomi | Mistral |
| Category | open | open |
| Modality | multimodal | multimodal |
| Release Date | Sep 2026 | Oct 2026 |
| Context Window | 1.0M | 1.0M |
| Parameters | 1.02T total, 42B active | 1.05T total, 49B active |
| License | MIT | Custom |
| Commercial use | free for commercial use | See license terms |
| Pricing per 1M tokens | ||
| Input $/1M | $0.435 | $0.700 |
| Output $/1M | $0.870 | $2.10 |
| Vision Tasks | ||
| Phrase Grounding | Not listed | Supported |
| Model Features | ||
Both support 3 more features | ||
Vision Evalsground-truth scores across 5 vision tasks, pooled at low effort | ||
| Overall | 50.3% | 56.8% |
| Avg cost / sample | $0.0011 | $0.0033 |
| Avg speed / sample | 9.45s | 31.23s |
| By task | ||
| Object Detection (low) | 42.0% ±1.1, Mean of 3 runs, range 40.9 to 43.1 | 59.3% ±0.7, Mean of 3 runs, range 58.5 to 60.0 |
| Object Detection (high) | 46.7% ±0.8, Mean of 3 runs, range 45.7 to 47.3 | 50.2% ±2.5, Mean of 3 runs, range 48.0 to 53.0 |
| Counting (low) | 50.0% ±2.0, Mean of 3 runs, range 48.6 to 52.7 | 54.5% ±0.7, Mean of 3 runs, range 54.0 to 55.4 |
| Counting (high) | 59.0% ±5.4, Mean of 3 runs, range 52.7 to 63.5 | 63.1% ±2.0, Mean of 3 runs, range 60.8 to 64.9 |
| Identification (low) | 76.0% ±1.6, Mean of 3 runs, range 75.0 to 78.1 | 85.4% ±3.1, Mean of 3 runs, range 81.3 to 87.5 |
| Identification (high) | 78.1% ±4.7, Mean of 3 runs, range 71.9 to 81.3 | 85.4% ±1.6, Mean of 3 runs, range 84.4 to 87.5 |
| OCR (low) | 48.6% | 46.1% |
| by category |
|
|
| OCR (high) | 49.9% | 54.5% |
| by category |
|
|
| Reasoning (low) | 35.1% ±2.6, Mean of 3 runs, range 32.5 to 37.8 | 38.9% ±0.3, Mean of 3 runs, range 38.4 to 39.1 |
| Reasoning (high) | 55.9% ±2.3, Mean of 3 runs, range 54.3 to 58.9 | 57.6% ±2.0, Mean of 3 runs, range 55.6 to 59.6 |
MiMo V2.6 Pro vs Mistral Large 4: Overview
MiMo V2.6 Pro is the flagship omni-modal foundation model in Xiaomi's MiMo V2.6 series, released as open weights alongside a Flash variant and a 9B distillation of Qwen3.5. It uses a sparse mixture-of-experts transformer with 1.02 trillion total parameters and roughly 42 billion activated per token, paired with a hybrid attention design that interleaves sliding-window and global attention layers to support a context window of about one million tokens. Dedicated encoders handle non-text inputs, including a vision encoder of roughly 681 million parameters and an audio tokenizer stack, so the model accepts text, images, video, and audio and returns text.
Post-training centers on large-scale reinforcement learning across thousands of interactive environments, combined with agentic grading, self-correction cold start, and a multi-prefix multi-teacher on-policy distillation stage that extends behavior to tasks that are hard to verify automatically.
Mistral Large 4, nicknamed Le Chonk, is a natively multimodal mixture-of-experts model from Mistral that accepts interleaved text and image input and produces text output. It uses a granular MoE design with roughly 1.05 trillion total parameters and 49 billion active per token, reported as 52 billion when embeddings and output layers are counted, paired with a 1.6 billion parameter vision encoder and a context window of one million tokens. The model is trained from scratch on about 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers and supports more than 160 languages. It behaves as a hybrid instruct and reasoning system, with a reasoning effort setting that selects between direct answers and longer deliberation, alongside function calling and structured output for agentic workflows.
Image understanding is a focus of this generation, covering documents, charts, technical drawings and natural scenes, and the model emits bounding box coordinates for visual grounding queries. Reported grounding results include 42 percent on Dense200 and 73 percent on the DIOR-RSVG remote sensing benchmark. Mistral describes agentic vision workflows in which the model zooms into gigapixel satellite imagery or engineering drawings to verify details, and reports coding results such as 62 percent on DeepSWE. Figures published at preview time are preliminary because the reinforcement learning phase is still in progress.