# MiniMax M2.5

> SWE-bench 80.2%. Top agentic coding.

Human version: https://kymaapi.com/models/minimax-m2.5
Live JSON: `GET https://kymaapi.com/v1/models` (no auth required)

## Facts

- **Model ID**: `minimax-m2.5` — pass this as `model` in the request body
- **Creator**: MiniMax
- **Released**: 2026-02-12
- **Context window**: 197K tokens
- **Max output**: 33K tokens per response — a hard ceiling, not a default
- **Price**: $0.3826 in / $1.346 out per 1M, cached input 10%
- **Capabilities**: tools, reasoning, caching

## Call it

```bash
curl https://kymaapi.com/v1/chat/completions \
  -H "Authorization: Bearer $KYMA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "minimax-m2.5", "messages": [{"role": "user", "content": "Hello"}]}'
```

## Positioning

MiniMax's top agentic coding model — 80.2% on SWE-bench — running in Kyma's frontier-open tier with a 100% success rate on production traffic over the measured period.

## About

MiniMax M2.5 is MiniMax's frontier-open model built for coding and agentic workflows. Its headline result is an 80.2% score on SWE-bench, the benchmark that measures whether a model can resolve real software issues end to end — which is why it's recommended for coding, agents, and debugging rather than general chat.

On Kyma it has earned real adoption: it ranks 11th of 62 models by tokens served, and coding agents — OpenClaw, Claude Code, and OpenCode — are among its top apps by traffic. Production requests in the measured window succeeded 100% of the time, every call gets automatic failover if a serving path degrades, and prompt caching bills repeated prompt prefixes at 10% of the input rate — which matters for agents that resend long system prompts on every turn.

Capability-wise it covers the full agent loop: a 196K-token context window, function calling, structured outputs, and extended reasoning, with up to 32K tokens of output per request.

## Use cases

- **Agentic coding** — Multi-step software tasks — read the codebase, plan, edit, verify — are exactly what its SWE-bench result measures.
- **Coding agent backbones** — OpenClaw, Claude Code, and OpenCode are already among its top traffic sources on Kyma.
- **Debugging sessions** — Extended reasoning plus a 196K context lets it hold the failing code, logs, and stack traces in one request while it works through the cause.
- **Structured tool pipelines** — Function calling and structured outputs keep multi-step agents emitting parseable, on-schema results.
- **Long-context code review** — The 196K window fits large diffs or several related modules in a single review pass.

## Not ideal for

Image inputs (it's text-only — pick a vision model) or latency-critical chat UX, where its medium-speed output (median 22 tokens/sec on Kyma) will feel slower than a fast-tier model.

## Pick something else when

- You want the absolute safest default → `qwen-3.6-plus`
- You want multimodal tool-heavy agents → `kimi-k2.6`
- You need faster cheap throughput → `qwen-3-32b`

## See also

- All models: https://kymaapi.com/models.md
- Pricing: https://kymaapi.com/pricing.md
- Other models by MiniMax: https://kymaapi.com/models?q=MiniMax