MiniMax-M3 Launch: 1M-Context Open Flagship at 50% Off

​

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)

​

Background

​

Deep Dive

​

Core Features

​

Benchmark Highlights

​

Technical Specs

​

Practical Usage

​

Recommended Scenarios

​

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)

​

Stack with Recharge Promotions

​

Summary & Recommendations

• ✅ 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

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)