Skip to main content

​
Key Takeaways

  • New open-weight flagship: MiniMax-M3 is now live on APIYI (mind the capitalization) — the first open-weight model to combine frontier coding-agent performance, a 1M-token context window, and native multimodality
  • Beats closed flagships on coding: SWE-Bench Pro 59.0, ahead of GPT-5.5 and Gemini 3.1 Pro; Terminal-Bench 2.1 at 66.0, MCP Atlas at 74.2
  • Tops autonomous browsing: BrowseComp 83.5, above Claude Opus 4.7 (79.3); first place on the Claw-Eval end-to-end agent benchmark
  • MSA sparse attention: MiniMax Sparse Attention replaces full attention with KV-block selection — 1M-context inference at roughly 1/20 the cost of the previous generation
  • Limited-time 50% off: APIYI matches the official discount — $0.30 input / $1.20 output per 1M tokens (0-512K tier), ending June 8, 2026, 00:00 (UTC+8)
  • Stack recharge bonuses: combined with APIYI recharge promotions, the effective price drops to roughly 41% of list (50% ÷ 1.2)
MiniMax officially released M3 on June 1, 2026, pledging open weights and a technical report on Hugging Face and GitHub within 10 days. Sources: minimax.io/blog/minimax-m3, venturebeat.com, openrouter.ai/minimax/minimax-m3. Data retrieved: 2026-06-05.

​
Background

MiniMax’s M series has always pushed the “long context + high cost-efficiency” frontier, and M3 takes that strategy to a new level. Released on June 1, 2026, M3 is positioned as the first open-weight model to combine frontier coding-agent capability, a 1M-token context window, and native multimodality — including image and video input plus desktop computer operation. The headline innovation is MSA (MiniMax Sparse Attention): replacing full attention with KV-block selection to drastically cut per-token compute at long context. Per official figures, inference at 1M tokens costs roughly 1/20 of the previous generation, with substantially faster prefill and decode. For the first time, “million-token context” is genuinely affordable. The most actionable part for developers is the price: MiniMax launched the model with a limited-time 50% discount, and APIYI has matched it — stackable with recharge promotions for an even lower effective rate.

​
Deep Dive

​
Core Features

1M-Token Context

Native million-token windowTrue 1M context powered by MSA sparse attention — entire repos, long video scripts, and extended agent trajectories in a single pass.

MSA Sparse Attention

Long-context cost cut to 1/20KV-block selection replaces full attention; officially 4×+ faster prefill/decode than open-source alternatives at long context.

Frontier Coding Agent

SWE-Bench Pro 59.0Ahead of GPT-5.5 and Gemini 3.1 Pro, Terminal-Bench 2.1 at 66.0 — production-ready for real software engineering tasks.

Native Multimodality

Image / video / computer useTrained on interleaved text-image data from inception; supports image and video input plus desktop operation, beating Gemini 3.1 Pro on OmniDocBench.

​
Benchmark Highlights

Data from the official MiniMax release and third-party coverage:
BenchmarkMiniMax-M3Comparison
SWE-Bench Pro (real-world SWE)59.0Above GPT-5.5, Gemini 3.1 Pro
Terminal-Bench 2.1 (terminal agent)66.0Leading among open models
MCP Atlas (tool use)74.2—
BrowseComp (autonomous browsing)83.5Above Claude Opus 4.7 (79.3)
Claw-Eval (end-to-end agent)#1Top of all tested models
OmniDocBench (document multimodal)LeadingAbove Gemini 3.1 Pro
In official demos, M3 autonomously reproduced academic papers and optimized CUDA kernels — first-tier long-horizon reasoning and engineering execution. With open weights landing within 10 days, in-house agent teams can run an “API + offline evaluation” dual track.

​
Technical Specs

Engineering parameters

  • Model ID: MiniMax-M3 (case-sensitive)
  • Attention: MSA (MiniMax Sparse Attention, KV-block selection)
  • Context window: 1M tokens (native)
  • Multimodality: ✅ image / video input, desktop computer operation
  • Billing: tiered by input length (0-512K / above 512K)
  • API compatibility: OpenAI ChatCompletions compatible
  • Open weights: weights + technical report on Hugging Face / GitHub within 10 days
M3 uses tiered billing: once a request’s input exceeds 512K tokens, the excess is billed at the higher tier ($0.60 input / $2.40 output per 1M tokens, limited-time rate). Budget accordingly for million-token workloads.

​
Practical Usage

Whole-Repo Coding Agent

1M context + SWE-Bench Pro 59.0 — large-repo refactors and cross-file PR tasks without chunking or retrieval

Autonomous Browsing & Research

Tops BrowseComp at 83.5 — ideal for deep-research agents and automated information gathering

Long Document / Video Understanding

Native multimodality + million-token context — analyze massive contracts, reports, or video content in one pass

Computer-Use Agents

Native desktop operation — build RPA, automated testing, and computer-use agents

​
Quick Start (OpenAI-Compatible API)

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-apiyi-key",
    base_url="https://api.apiyi.com/v1"
)

# Note the capitalization: MiniMax-M3
resp = client.chat.completions.create(
    model="MiniMax-M3",
    messages=[
        {"role": "system", "content": "You are a senior full-stack engineer specializing in PR-level tasks in large codebases."},
        {"role": "user", "content": "Read through this entire repo, find the rate-limiting middleware bottleneck, and propose a fix"}
    ],
    temperature=0.3,
)
print(resp.choices[0].message.content)

curl https://api.apiyi.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-your-apiyi-key" \
  -d '{
    "model": "MiniMax-M3",
    "messages": [{"role": "user", "content": "Explain MSA sparse attention in one paragraph"}]
  }'

​
Best Practices

  • Model ID capitalization: the ID is MiniMax-M3 — a wrong case returns a 404 model-not-found error
  • Cost control: keep single-request input under 512K tokens to stay on the lowest tier; summarize/trim for ultra-long tasks
  • Temperature: temperature=0.2 ~ 0.4 recommended for agent/coding workloads
  • Streaming: prefill takes longer at extreme context — enable streaming for better perceived latency
  • Discount window: the 50% off rate ends June 8, 00:00 (UTC+8) — schedule heavy evaluations and batch jobs before then

​
Pricing & Availability

​
Price Table (USD / 1M tokens, limited-time 50% off)

Input length tierPrompt (input)Completion (output)
0 - 512K$0.3000$1.2000
Above 512K$0.6000$2.4000
APIYI matches MiniMax’s official limited-time 50% discount — the table above is the current live rate, billed in tiers by input length. The discount ends June 8, 2026, 00:00 (UTC+8); subsequent pricing is to be determined.

​
Stack with Recharge Promotions

The 50%-off rate stacks with APIYI recharge bonuses, bringing the effective price down to roughly 41% of list (50% ÷ 1.2):

Recharge Promotions

See the latest recharge bonus tiers — larger top-ups earn bigger bonuses

​
Summary & Recommendations

MiniMax-M3 is the first model to pack “open weights + million-token context + multimodal agent” into a single package:
  • ✅ Benchmark wins: SWE-Bench Pro 59.0 above GPT-5.5 / Gemini 3.1 Pro; BrowseComp 83.5 above Opus 4.7
  • ✅ Million-token context, actually affordable: MSA sparse attention cuts 1M-context cost to 1/20 of the previous generation
  • ✅ Pricing window: limited-time 50% off + recharge bonuses ≈ 41% of list, ending June 8, 00:00 (UTC+8)
  • ✅ Open weights incoming: weights and technical report within 10 days — start with the API now, run offline evals later
Recommended next steps:
  1. Add MiniMax-M3 to your A/B rotation in Claude Code / Cursor / in-house agents, focusing on whole-repo and long-horizon tasks
  2. Run a full evaluation of long-document / video-understanding workloads during the 50%-off window
  3. Stack recharge bonuses to push the effective rate to ~41%, and schedule batch jobs before June 8, 00:00 (UTC+8)
Sources & dates
  • Official MiniMax release: minimax.io/blog/minimax-m3
  • Third-party coverage: venturebeat.com, techtimes.com, officechai.com
  • Pricing page: openrouter.ai/minimax/minimax-m3
  • Data retrieved: 2026-06-05