TheRouter.aiTheRouter.ai
Back to Models

MiniMax M2.7 Highspeed

minimaxminimax/m2.7-highspeed

MiniMax M2.7 Highspeed variant with ~100 tokens/sec output speed. Same capabilities as M2.7 at 2x cost for latency-sensitive applications.

MiniMax M2.7 announcement — minimax.io ↗Vercel AI Gateway changelog — M2.7 Highspeed ↗GitHub repository (M2.7 standard weights) ↗Hugging Face model card (M2.7 standard) ↗MiniMax Open Platform — API docs ↗

MiniMax M2.7 Highspeed is the high-throughput variant of the M2.7 model, tuned for ~100 tokens/sec output — roughly 1.6× faster than the standard M2.7 and 3× faster than Claude Opus 4.6. It delivers the identical same agentic coding, tool orchestration, and office automation capabilities as standard M2.7 at double the price ($0.80/M input, $3.20/M output). Built on the same MoE (Mixture of Experts) architecture with interleaved thinking chains, it targets latency-sensitive production workloads where response time matters more than cost.

For TheRouter operators routing latency-sensitive traffic — real-time coding assistants, voice agent backends, interactive chat, or agentic loops that need fast token generation — M2.7 Highspeed is the go-to variant within the MiniMax M2-series. Standard M2.7 remains the cost-optimized choice at half the price when speed isn't critical. The model's benchmarks are identical to standard M2.7: SWE-Pro 56.22%, Terminal Bench 2 57.0%, VIBE-Pro 55.6%, and the highest ELO among open-source models on GDPval-AA (1495).

Best for
  • • Latency-sensitive coding assistants — real-time code completion, debugging, and refactoring where sub-second response perception matters
  • • Voice agent backends — voice-to-voice agent pipelines needing fast text generation between ASR and TTS stages
  • • Interactive multi-turn chat — production chatbots where users expect rapid conversational pace and low perceived latency
  • • Agentic loop acceleration — when agent self-improvement loops (MLE Bench Lite 66.6% medal rate) need faster eval cycles
  • • Office automation with real-time feedback — complex Excel/PPT/Word editing that benefits from faster iteration of multi-round revisions
Reach for something else if
  • • Cost-sensitive bulk inference — standard M2.7 is half the price at the same capability. Use minimax/m2.7 for cost-optimized batch processing
  • • Multimodal workloads — M2.7 Highspeed is text-in/text-out only. For vision, audio, or image gen, route to Doubao Seed 1.6 Vision or GPT-Image
  • • Extremely long-context retrieval — at 204K context, the Highspeed variant prioritizes speed over memory depth. For very large document analysis, consider Gemini 2.5 Pro with 1M+ context
Context Length
205K
Max Output
66K
Input Priceper 1M tokens
$0.648/ 1M tokens
Output Priceper 1M tokens
$2.59/ 1M tokens

Modalities

text→text

Pricing Breakdown

TypeRate
Input$0.648 / 1M tokens
Output$2.59 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Variant typeHigh-throughput accelerated variant (~100 tokens/sec output)vercel.com/changelog ↗verified
Relative speed vs standard~1.6× faster output than M2.7 standard; ~3× faster than Claude Opus 4.6minimax.io/news ↗to verify
Release date2026-04-06 (alongside standard M2.7)www.minimax.io ↗verified
ArchitectureSame MoE (Mixture of Experts) as M2.7; accelerated inference path via optimized batching/infrastructurewww.minimax.io ↗verified
LicenseHighspeed is an API-only variant (not open-weight); standard M2.7 weights available under Modified-MITgithub.com ↗verified
Self-evolution capabilitySame as M2.7 — recursive improvement of agent harness and RL pipeline via self-feedback loopswww.minimax.io ↗verified
Skill adherence97% across 40+ complex skills (each >2,000 tokens)www.minimax.io ↗verified
Predecessor variantMiniMax M2.5 Highspeed (minimax/m2.5-highspeed)vercel.com ↗verified

Benchmarks

BenchmarkDistributionScoreSource
SWE-Pro
Same as standard M2.7 — matches GPT-5.3-Codex
56.22%accuracywww.minimax.io ↗
VIBE-Pro
Repo-level full-project code gen — near Opus 4.6
55.6%accuracywww.minimax.io ↗
Terminal Bench 2
Complex engineering system comprehension
57.0%accuracywww.minimax.io ↗
GDPval-AA
Highest ELO among open-source models
1495ELOwww.minimax.io ↗
MLE Bench Lite
Self-evolution capability — 3-trial average
66.6%medal ratewww.minimax.io ↗
SWE Multilingual
Multi-language bug fix benchmark
76.5www.minimax.io ↗
Multi-SWE-Bench
Cross-task SWE benchmark
52.7www.minimax.io ↗
NL2Repo
Natural language to repository-level code
39.8%accuracywww.minimax.io ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "minimax/m2.7-highspeed",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

Basic text generation via chat completions endpoint. Same interface as standard M2.7 but with accelerated token generation. Supports temperature, max_tokens, top_p, stop, and response_format (JSON mode).

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"minimax/m2.7-highspeed","messages":[{"role":"user","content":"Write a Rust function to parse CSV"}],"temperature":0.7,"max_tokens":4096}'

More from minimax

minimax/m2.7Standard M2.7 — same capabilities at half the priceminimax/m2.5-highspeedM2.5 high-throughput variant — predecessor highspeedminimax/m2.5Standard M2.5 — predecessor to M2.7minimax/m2.1Earlier M2-series open-weight model (Dec 2025)

Similar models

Cross-provider sibling models
deepseek/deepseek-v3.2
Open-weight MoE rival with comparable coding strength
qwen/qwen3-235b
Alibaba Qwen3 MoE — comparable scale and capability
cohere/command-a
Open-weight tool-use specialist
anthropic/claude-haiku-4.5
Fast, cost-efficient Claude — comparable speed tier
google/gemini-2.5-flash
Google's fast multimodal model at similar price point

News & changes

2026-04-06

MiniMax M2.7 launched alongside Highspeed variant

MiniMax announced the M2.7 model family, including a Highspeed variant targeting ~100 tokens/sec output at 2x the standard pricing. The Highspeed variant is promoted by Vercel AI Gateway for latency-sensitive use cases like real-time coding assistants and voice agent backends.

re-authored by TheRoutervercel.com/changelog ↗

Frequently asked

What is the difference between M2.7 standard and Highspeed?

The Highspeed variant delivers identical agentic coding, tool orchestration, and office automation capabilities as standard M2.7, but at ~100 tokens/sec output — roughly 1.6x faster. The trade-off is 2x pricing ($0.80/M input, $3.20/M output vs $0.40/$1.60). Use Highspeed when latency matters, standard when cost matters.

vercel.com ↗
Is M2.7 Highspeed available as open-weight? Can I self-host it?

No. M2.7 Highspeed is an API-only variant. Only the standard M2.7 weights are released under Modified-MIT license for self-hosting. The Highspeed variant requires the provider's accelerated inference infrastructure and is only available via API endpoints.

github.com ↗
How fast is M2.7 Highspeed compared to other models?

MiniMax claims ~100 tokens/sec for Highspeed vs ~60 tokens/sec for standard M2.7 — roughly 3x faster than Claude Opus 4.6. Independent third-party testing (Thomas Wiegold blog) confirms approximately 100 TPS for the Highspeed variant. For comparison, GPT-5-mini and Claude Haiku 4.5 operate in a similar speed tier (~80-120 TPS).

thomas-wiegold.com ↗
Which models are in the MiniMax M2-series on TheRouter?

TheRouter currently lists M2, M2.1, M2.1 Highspeed, M2.5, M2.5 Highspeed, M2.7, and M2.7 Highspeed — covering the full M2-series. The family progresses from M2 (base) through M2.1 (Dec 2025, coding-focused), M2.5 (Feb 2026, agent-native), to M2.7 (Apr 2026, self-evolving). Each model also has a Highspeed variant at 2x cost.

What API should I use on TheRouter to call M2.7 Highspeed?

Use the OpenAI-compatible /v1/chat/completions endpoint with model="minimax/m2.7-highspeed". The same SDK works for both standard and Highspeed — just change the model id. Authentication uses your TheRouter API key via the Authorization header.

Fact ledger — every claim on this page traces here
sourceURLretrieved
Variant typevercel.com/changelog ↗2026-05-27verified
Relative speed vs standardminimax.io/news ↗2026-05-27to verify
Release datewww.minimax.io ↗2026-05-27verified
Architecturewww.minimax.io ↗2026-05-27verified
Licensegithub.com ↗2026-05-27verified
Self-evolution capabilitywww.minimax.io ↗2026-05-27verified
Skill adherencewww.minimax.io ↗2026-05-27verified
Predecessor variantvercel.com ↗2026-05-27verified
SWE-Prowww.minimax.io ↗2026-05-27verified
VIBE-Prowww.minimax.io ↗2026-05-27verified
Terminal Bench 2www.minimax.io ↗2026-05-27verified
GDPval-AAwww.minimax.io ↗2026-05-27verified
MLE Bench Litewww.minimax.io ↗2026-05-27verified
SWE Multilingualwww.minimax.io ↗2026-05-27verified
Multi-SWE-Benchwww.minimax.io ↗2026-05-27verified
NL2Repowww.minimax.io ↗2026-05-27verified
MiniMax M2.7 launched alongside Highspeed variantvercel.com/changelog ↗2026-05-27verified
What is the difference between M2.7 standard and Highspeed?vercel.com ↗2026-05-27to verify
Is M2.7 Highspeed available as open-weight? Can I self-host it?github.com ↗2026-05-27to verify
How fast is M2.7 Highspeed compared to other models?thomas-wiegold.com ↗2026-05-27to verify
Help & contact