MiniMax M2.5 Highspeed variant with ~100 tokens/sec output speed. Same capabilities as M2.5 at 2x cost for latency-sensitive applications.
MiniMax M2.5-Highspeed (also referred to as the Lightning variant) is the high-throughput version of MiniMax M2.5, released February 12, 2026. It shares the identical 230B-parameter MoE architecture (10B active per token) and the same Forge RL training regimen as the standard M2.5, but is served at a steady ~100 tokens per second — roughly double the standard variant's 50 TPS. On TheRouter it is priced at 2x the base M2.5 rate ($0.80/M input, $3.20/M output), exactly mirroring the official MiniMax pricing ratio between M2.5-Lightning and M2.5.
This variant is the right choice when end-to-end latency is the binding constraint — real-time agent interactions, user-facing chatbots, or throughput-sensitive batch workloads where a 100 TPS output rate meaningfully impacts user experience. For cost-optimized or sustained agent operations, the standard M2.5 variant (minimax/m2.5) at 50 TPS and half the price is the better default. Capability-wise both variants are identical: same 80.2% SWE-Bench Verified, same BrowseComp 76.3%, same tool-calling behavior. The only difference is throughput and cost.
| Type | Rate |
|---|---|
| Input | $0.648 / 1M tokens |
| Output | $2.59 / 1M tokens |
| Release date | February 12, 2026www.minimax.io ↗ | verified |
| Variant type | Highspeed / Lightning — capability-identical to M2.5 at 2x throughput (100 TPS) and 2x costwww.minimax.io ↗ | verified |
| Architecture | Mixture of Experts (MoE) — 230B total / 10B activated per tokenwww.minimax.io ↗ | verified |
| Output throughput | ~100 tokens per second (steady)www.minimax.io ↗ | verified |
| License | Modified MIT — commercial use allowed with attribution (display 'MiniMax M2.5' on UI)github.com ↗ | verified |
| RL framework | Forge — agent-native RL framework with CISPO algorithmwww.minimax.io ↗ | verified |
| Training environments | 200,000+ real-world simulated environmentswww.minimax.io ↗ | verified |
| Training languages | Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Rubywww.minimax.io ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified Same score as standard M2.5 — capability-identical. On Droid scaffold: 79.7% (vs Opus 4.6 78.9%), on OpenCode scaffold: 76.1% (vs Opus 4.6 75.9%). Highspeed variant completes tasks ~37% faster end-to-end than M2.1; the standard variant matches Opus 4.6 speed at 22.8 min/task, and the Highspeed variant is proportionally faster. | 80.2% | www.minimax.io ↗ | |
Multi-SWE-bench First place at release — multi-language software engineering benchmark. Same score as standard M2.5. | 51.3% | www.minimax.io ↗ | |
BrowseComp With context management enabled. Same score as standard M2.5 — capability-identical across all benchmarks. | 76.3% | www.minimax.io ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "minimax/m2.5-highspeed",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'M2.5-Highspeed uses the same OpenAI-compatible v1/chat/completions endpoint as any TheRouter model. Simply set the model field to minimax/m2.5-highspeed.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"minimax/m2.5-highspeed","messages":[{"role":"user","content":"Write a React component for a file uploader with drag-and-drop support."}]}'MiniMax released M2.5, an open-weight frontier model achieving 80.2% SWE-Bench Verified and 51.3% Multi-SWE-Bench. The model ships in two throughput tiers: standard (50 TPS) and Lightning/Highspeed (100 TPS). Both variants share the same architecture, benchmarks, and capabilities — only speed and pricing differ.
Both variants share the exact same architecture, weights, and capabilities — same SWE-Bench scores, same tool-calling behavior, same context length. The only difference is throughput: M2.5-Highspeed runs at ~100 tokens/sec at 2x the cost, while standard M2.5 runs at ~50 tokens/sec at half the price. Choose Highspeed when latency is the priority; choose standard when cost efficiency matters more.
MiniMax describes M2.5-Lightning as the same model in a faster serving tier. Both variants use identical weights and architecture — they are not independently trained models. On TheRouter they appear as separate slugs because each variant has its own pricing and throughput characteristics, but the model output is identical for the same input.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | www.minimax.io ↗ | 2026-05-27 | verified |
| Variant type | www.minimax.io ↗ | 2026-05-27 | verified |
| Architecture | www.minimax.io ↗ | 2026-05-27 | verified |
| Output throughput | www.minimax.io ↗ | 2026-05-27 | verified |
| License | github.com ↗ | 2026-05-27 | verified |
| RL framework | www.minimax.io ↗ | 2026-05-27 | verified |
| Training environments | www.minimax.io ↗ | 2026-05-27 | verified |
| Training languages | www.minimax.io ↗ | 2026-05-27 | verified |
| SWE-bench Verified | www.minimax.io ↗ | 2026-05-27 | verified |
| Multi-SWE-bench | www.minimax.io ↗ | 2026-05-27 | verified |
| BrowseComp | www.minimax.io ↗ | 2026-05-27 | verified |
| MiniMax M2.5 released with Lightning (Highspeed) variant | minimax.io/news ↗ | 2026-05-27 | verified |