MiniMax M2.1 Highspeed variant with ~100 tokens/sec output speed. Same capabilities as M2.1 at 2x cost for latency-sensitive applications.
MiniMax M2.1 Highspeed is an inference-optimized variant of the open-weight MiniMax M2.1 (230B total / 10B active MoE) that delivers approximately 100 tokens per second — roughly 1.6x the output speed of the standard M2.1 variant, at 2x the per-token cost. Critically, the Highspeed variant runs the same model weights, produces the same output quality, and offers the same 204,800-token context window as standard M2.1. The speed gain comes entirely from inference-layer optimizations: tuned MoE expert routing for lower latency, more aggressive batching under concurrency, and memory-management improvements for long-context requests.
For a TheRouter operator, M2.1 Highspeed is the latency-sensitive choice when you need M2.1's coding and agent capabilities but user-facing latency matters — live coding assistants, real-time chat, streaming agent output. It costs $0.80/M input tokens and $3.20/M output tokens via TheRouter (2x the $0.40/$1.60 of standard M2.1). Use standard M2.1 for batch jobs, offline evaluation, or any scenario where end-to-end latency is not user-visible; reach for Highspeed when the gap between ~60 tps and ~100 tps makes the difference between "this feels snappy" and "this feels like waiting."
| Type | Rate |
|---|---|
| Input | $0.648 / 1M tokens |
| Output | $2.59 / 1M tokens |
| Availability | Available as MiniMax-M2.1-highspeed on platform.minimax.io and via TheRouter at api.therouter.ai/v1 with model ID minimax/m2.1-highspeedplatform.minimax.io — Supported Models ↗ | verified |
| Output speed | ~100 tokens per second (official MiniMax figure); standard M2.1 is ~60 tps for comparisonplatform.minimax.io — Model Overview ↗ | verified |
| Architecture | 230B total / 10B active MoE — same weights as standard M2.1; only inference-layer differs (optimized routing, batching, KV memory)github.com/MiniMax-AI — M2.1 repository ↗ | verified |
| Context window | 204,800 tokens — identical to standard M2.1platform.minimax.io ↗ | verified |
| Prompt caching | Automatic prompt caching supported — reduces latency on cached tokens by 60–80% for repeated system prompts and long-context preamblesplatform.minimax.io — Model Overview ↗ | to verify |
| License | Same Modified-MIT as M2.1 (commercial use permitted for open weights); Highspeed API access through MiniMax platform or TheRouter is standard commercial API licensinghuggingface.co/MiniMaxAI ↗ | to verify |
| Supported inference backends | SGLang, vLLM, Transformers, MLX-LM, KTransformers (shared with standard M2.1 open weights)github.com/MiniMax-AI ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified Consistent across multiple frameworks (Claude Code, Droid, mini-SWE-agent) — framework-agnostic coding performance | 74.0%% | minimax.io — official announcement ↗ | |
Multi-SWE-bench | 49.4%% | minimax.io — official announcement ↗ | |
SWE-bench Multilingual Rust, Java, Golang, C++, Kotlin, TypeScript, JavaScript, Objective-C | 72.5%% | minimax.io — official announcement ↗ | |
VIBE (aggregate) Web 91.5 | Android 89.7 | iOS 88.0 | Simulation 87.1 | Backend 86.7 — end-to-end zero-to-one app scaffolding | 88.6%% | minimax.io — official announcement ↗ | |
Toolathlon | 43.5%% | minimax.io — official announcement ↗ | |
BrowseComp | 47.4%% | minimax.io — official announcement ↗ | |
AIME 2025 | 83.0%% | minimax.io — official announcement ↗ | |
MMLU | 90.0%% | minimax.io — official announcement ↗ |
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/m2.1-highspeed",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat completion via TheRouter — only baseURL and model change from the official OpenAI SDK.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax/m2.1-highspeed",
"messages": [
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write a Rust function to parse a CSV line."}
],
"temperature": 0.3,
"max_tokens": 4096
}'MiniMax launched M2.1 (230B total / 10B active MoE) with strong SWE-bench Verified (74.0%), multilingual programming support, and the new Highspeed API tier offering ~100 tps for latency-sensitive applications. The Highspeed variant uses identical model weights with inference-layer optimizations.
No. Highspeed runs the exact same model weights (230B total / 10B active MoE) and produces identical output quality. The speed gain comes entirely from inference-layer optimizations: tuned MoE expert routing, concurrency batching, and KV-cache memory management. Same answers, faster.
Use Highspeed when end-to-end latency is user-facing — live coding assistants, real-time chatbots, streaming agent output. Use standard M2.1 (minimax/m2.1) for batch jobs, offline evaluation, or any scenario where the ~40 tps speed gap is not user-visible. The 2x per-token cost of Highspeed ($0.80/$3.20 vs $0.40/$1.60) makes this trade-off strictly about latency vs cost.
No. The Highspeed variant is an API-tier SKU on MiniMax's platform (and via TheRouter). The underlying model weights are identical to the open-weight M2.1 release (Modified-MIT licensed), but the inference optimizations that produce the speed gain are a proprietary API feature. For self-hosting, use the standard M2.1 open weights which achieve ~60-70 tps on consumer GPUs.
| source | URL | retrieved | |
|---|---|---|---|
| Availability | platform.minimax.io — Supported Models ↗ | 2026-05-27 | verified |
| Output speed | platform.minimax.io — Model Overview ↗ | 2026-05-27 | verified |
| Architecture | github.com/MiniMax-AI — M2.1 repository ↗ | 2026-05-27 | verified |
| Context window | platform.minimax.io ↗ | 2026-05-27 | verified |
| Prompt caching | platform.minimax.io — Model Overview ↗ | 2026-05-27 | to verify |
| License | huggingface.co/MiniMaxAI ↗ | 2026-05-27 | to verify |
| Supported inference backends | github.com/MiniMax-AI ↗ | 2026-05-27 | verified |
| SWE-bench Verified | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| Multi-SWE-bench | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| SWE-bench Multilingual | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| VIBE (aggregate) | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| Toolathlon | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| BrowseComp | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| AIME 2025 | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| MMLU | minimax.io — official announcement ↗ | 2026-05-27 | to verify |
| MiniMax releases M2.1 with coding focus and Highspeed variant | minimax.io — official announcement ↗ | 2026-05-27 | verified |
| Does the Highspeed variant produce different output quality vs standard M2.1? | AIMLAPI — MiniMax HighSpeed guide ↗ | 2026-05-27 | to verify |