Coding god,
everything else mid.
MiniMax M2.5 (released February 2026): 229B MoE, open-weight. The best coding model in the world can't do basic science.
Coding
#1 open-weight · within 0.6% of Opus
Science
44 points behind Opus · below frontier
The numbers
World-class at one thing, below average at everything else.
| Benchmark | MiniMax M2.5 | Opus 4.6 | Kimi K2.5 | Notes |
|---|---|---|---|---|
| SWE-bench Verified (coding) | 80.2% | 80.8% | 76.8% | MiniMax within 0.6% of Opus |
| GPQA Diamond (science) | 47% | 91.3% | 87.6% | 44 points behind Opus |
| AIME 2026 (math) | ~78-88% | 92.8% | 96.1% | Kimi is the math king |
| Arena ELO (overall) | none | 1503 | 1447 | No published ELO for MiniMax |
The GPQA row is the story. 47% on science-and-reasoning when the frontier is 91%. Kimi K2.5 scores 87.6% on GPQA with thinking mode turned off, 40 points higher than MiniMax with thinking always on. MiniMax didn't just deprioritize science. It abandoned it.
How they did it
Specialization is a training choice.
MiniMax M2.5 uses interleaved thinking: multi-step reasoning baked into the architecture, always on, no off switch. Hours spent optimizing for SWE-bench-style tasks (cloning repos, navigating file trees, writing patches, running tests) are hours not spent on scientific reasoning or mathematical proof chains.
The 47% GPQA is the price of 80.2% SWE-bench. MiniMax chose its battle and won it.
Total parameters
MoE: routes to specialized experts
Active per token
Tuned for code generation patterns
Hardware
What it costs to run.
MiniMax is cheaper than Kimi but only gives you coding. Kimi gives you everything.
| Model | Architecture | VRAM | Hardware | Cost |
|---|---|---|---|---|
| MiniMax M2.5 | 229B MoE, 10B active | ~457 GB | 4× H100 | $100-120K buy / ~$5K/mo |
| Kimi K2.5 | ~1T MoE, 32B active | ~600 GB | 8× H100 | $180-240K buy / ~$8K/mo |
| Claude Opus 4.6 | Unknown (API only) | API only | API only | ~$15/$75 per MTok |
Decision matrix
When MiniMax M2.5 is the right choice.
| Use case | MiniMax? | Why |
|---|---|---|
| Pure coding agent | Yes | If all it does is read code, write patches, and run tests: MiniMax is nearly Opus-level |
| Code + architecture reasoning | Maybe | Depends how much reasoning the architecture decisions require |
| Code + data analysis | No | Data analysis requires math and statistical reasoning, MiniMax’s weak point |
| General-purpose assistant | No | 47% GPQA means it can’t handle basic science questions |
| Research + coding | No | Any task that requires understanding papers, specs, or protocols is out |
Open weights reached 80.2% SWE-bench. The model ceiling is reachable.
The next frontier is the scaffold, not the model.The bigger picture
Specialist models are a signal.
Before MiniMax M2.5, the assumption was that open-weight models couldn't touch frontier APIs on coding. Now the gap is 0.6%. That changes the conversation from "which model is best?" to "what's wrapped around the model?"
SWE-bench Pro shows a 22+ point swing between different agent scaffolds using the same model. That's 3× the gap between MiniMax and Opus. The retrieval infrastructure, context management, and tool orchestration matter more than the model weights.
MiniMax proved the model ceiling is reachable. The next frontier is the scaffold, and that's where Interlinked Files lives. Machine-wide retrieval that turns any model into a better coding agent.