Text and chat · MiniMax
MiniMax
MiniMax's open models, tuned for coding and agent work, in private mode at a low price.
Private Low costWhat sets it apart
MiniMax builds mixture-of-experts models aimed at coding and multi-step agent work, with only about 10B parameters active in the M2 line. M3 moves to 1M context and reads images and video, at a price that stays low.
- Only 10B active of 230B in M2.5 and M2.7
- Tuned for coding and agent loops
- M3: 1M context, text, image and video
Strengths and limits
Where it shines
- Built for coding and multi-step agent tasks
- Low price, private mode, tool calling
Think twice when
- M2.5 and M2.7 read text only
- M3 is a preview and may change
- M2.7: 198K context
- M3 preview: 512K context, vision
Use cases
Fix a bug across several files
Here are three files from my checkout flow and the error log. Find why the total sometimes shows twice the price, then propose a minimal fix.
The M2 line is tuned for coding and tool-driven repair loops at a low price.
Read a whole long document with figures
Read this 200-page report with its charts and list every figure that contradicts the executive summary.
M3 accepts images and a 1M token context window.
Against its rivals
- GLMGLM is the closest open coding and agent family, with larger flagships.#13 of 90 on Artificial Analysis Intelligence Index (45 points)
- KimiKimi targets the same agent work and generally costs more.#15 of 90 on Artificial Analysis Intelligence Index (44 points)
- DeepSeekDeepSeek competes on price and coding, with no image input in most versions.#24 of 90 on Artificial Analysis Intelligence Index (39 points)
Version history
Each version, dated, with what it changed over the one before.
Hollow points are earlier versions no longer offered in Sub Rosa.
See the numbers
| Version | Released | points |
|---|---|---|
| MiniMax M2.5 | 2026-02 | 23 |
| MiniMax M2.7 | 2026-03-18 | 23 |
- MiniMax M2.5
- 80.2% on SWE-bench Verified, per MiniMax, with stronger tool use and search
- Plans the architecture of a project before writing code
- MiniMax M2.7
- Self-evolution: an internal version optimized its own coding scaffold over 100+ rounds
- Agent Teams for multi-agent collaboration and broader tool calling
- MiniMax M3Latest
- 428B total and 23B active parameters with 1M context
- Native multimodal input (text, image, video) and sparse attention for 9x faster prefill, per MiniMax
Benchmarks
How to read these scores: what a benchmark tells you.
Artificial Analysis Intelligence Index: #61 of 90 models measured in the catalog. One score averaging about ten independent tests of reasoning, knowledge, math, coding and agent work, all run the same way for every model. Higher is better. Each model is listed at its highest reasoning effort. Two or three points apart is a close call; ten points is a different class of model.
See the numbers
| Model | points |
|---|---|
| 1. Claude Opus 5.5 | 58 |
| 2. Claude Sonnet 5.5 | 56 |
| 3. Claude Fable 5.1 | 53 |
| 4. GPT-6 Astra | 53 |
| 5. Gemini 4 Argon | 53 |
| 6. GPT-6.1 Sol | 52 |
| 7. Muse Wonder | 52 |
| 8. Claude Opus 5 | 51 |
| 9. Claude Fable 5 | 50 |
| 10. GPT-6 Sol | 48 |
| 61. MiniMax M2.7 | 23 |
| 62. MiniMax M2.5 | 23 |
Announced by the maker
Measured by the maker with its own settings: useful, but not directly comparable with the independent scores above.
| Benchmark | Model | Score | Source |
|---|---|---|---|
| swe-bench-verified | MiniMax M2.5 | 80.2 | huggingface.co |
All 3 models in this family
| Model | Use | Privacy | Price |
|---|---|---|---|
MiniMax M2.7minimax-m27 | Text | Private | $0.674 per million tokens written$0.000472 a page written |
MiniMax M2.5minimax-m25 | Text | Private | $0.427 per million tokens written$0.000299 a page written |
MiniMax M3 Previewminimax-m3-preview | Text | Private | $0.54 per million tokens written$0.000378 a page written |
Specifications
| Specification | MiniMax M3 (preview) | MiniMax M2.7 | MiniMax M2.5 |
|---|---|---|---|
| Released | June 1, 2026 | March 18, 2026 | February 2026 |
| Parameters | 428B total, 23B active | 230B total, 10B active | 230B total, 10B active |
| Weights | Open, MiniMax Community License | Open, Non-commercial | Open, Modified MIT |
| Accepts | text, images, video | text | text |
| Produces | text | text | text |
| Context | 524K tokens | 198K tokens | 198K tokens |
| Longest answer | 66K tokens | 33K tokens | 33K tokens |
| Abilities | Reads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for code | Uses Sub Rosa’s tools, Thinks before answering, Tuned for code | Uses Sub Rosa’s tools, Thinks before answering, Tuned for code |
| Price | $0.54 per million tokens written | $0.674 per million tokens written | $0.427 per million tokens written |
| Privacy | Private | Private | Private |
Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.
Sources
Everything on this page comes from these pages, read on October 16, 2026 or as dated on each score.