MiniMax M3.1 Flash API

Space Bunny was M3.1 Flash running without a name

MiniMax published MiniMax M3.1 Flash on October 19, 2026 as its next-generation Flash model for agentic tasks, coding, computer use, and cybersecurity. In the same release post MiniMax confirmed that Space Bunny was an early version of MiniMax M3.1 Flash, running anonymously while it handled roughly 200 trillion tokens.

MiniMax M3.1 Flash benchmarks

MiniMax released M3.1 Flash with one comparison table: 15 benchmarks covering agentic office work, computer use, coding, cybersecurity and general capability, scored against DeepSeek V4.1 Flash-1017, DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. The table below reproduces it in full. All figures are vendor-reported, so use them to build a shortlist and run a representative test before routing production traffic.

The results are uneven across row groups. M3.1 Flash has the best score in three rows, all agentic office work: Automation-Bench v1.0.6 (62.8), LOCA-Bench 256K (60.3) and BankerToolBench, where it ties Claude Opus 5 at 76.2. It trails on terminal work and repository-scale coding: Terminal Bench 4.0 is 38.1 against 42.3 for DeepSeek V4.1 Flash-1017 and 51.8 for Opus 5, and NL2Repo-Bench is 61.2 against 68.2 and 75.3 for the same two models.

Scores by reasoning effort

MiniMax also charted M3.1 Flash at each reasoning-effort setting. The comparison-table scores for Automation-Bench and DeepSWE (62.8 and 73.7) are the max setting, and on Automation-Bench every setting cost less than $0.05 per task.

Third-party coverage has been less uniform. VentureBeat reported an Intelligence Index score of 42 at maximum reasoning effort against 29 for M3, and also flagged a high hallucination rate, while a Hacker News discussion questioned how well the model's benchmark position survives real use. Both are reasons to evaluate on your own prompts rather than on a leaderboard position.

MiniMax M3.1 Flash pricing

MiniMax M3.1 Flash is free on TokenRa until October 31, 2026: $0 per 1M input, output, and cached input tokens, with no platform subscription. For reference, MiniMax lists $0.15 per 1M input tokens, $0.60 per 1M output tokens, and $0.015 per 1M cached input tokens, and states that overall usage cost is about 70% lower than its predecessor M3.

How to call MiniMax M3.1 Flash through TokenRa

Create a TokenRa key, confirm that minimax-m3.1-flash is enabled for your account, and send an OpenAI-compatible request. Existing OpenAI SDK clients need only a different base URL and model identifier.

curl https://tokenra.io/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"minimax-m3.1-flash","messages":[{"role":"user","content":"Write a Python function that retries a flaky HTTP call with backoff, then explain the tradeoffs."}]}'

from openai import OpenAI

client = OpenAI(
    base_url="https://tokenra.io/v1",
    api_key="YOUR_API_KEY",
)

resp = client.chat.completions.create(
    model="minimax-m3.1-flash",
    messages=[
        {"role": "system", "content": "You are a careful coding assistant."},
        {"role": "user", "content": "Refactor this function and list the risks."},
    ],
)
print(resp.choices[0].message.content)

MiniMax M3.1 Flash model specifications

As of October 22, 2026, MiniMax has not published a parameter count for M3.1 Flash. Its predecessor, MiniMax-M3, is a mixture-of-experts model with about 428B total and 23B activated parameters, according to the MiniMax-M3 model card; some press coverage quotes the same figures for M3.1 Flash.

Space Bunny and MiniMax M3.1 Flash: one model, two names

Space Bunny and MiniMax M3.1 Flash name the same model at different stages. MiniMax confirmed that Space Bunny was an early version of the released model, so new integrations should use the released identifier. The anonymous stealth build keeps its own guide: Space Bunny Alpha (stealth version).

MiniMax M3.1 Flash vs DeepSeek V4.1 Flash

Both are Flash-tier models with a 1M-token context, aimed at cost-sensitive agent and coding work, and both run on the same TokenRa key. DeepSeek's current Flash API model is DeepSeek V4.1 Flash 1017, released October 17, 2026. DeepSeek has retired V4 Flash, and requests to the old deepseek-v4-flash name now go to the 1017 build. On the six benchmarks where MiniMax's table scores both models, DeepSeek V4.1 Flash-1017 leads five and M3.1 Flash leads Automation-Bench. M3.1 Flash wins on price: its list rate is half of DeepSeek's peak rate, and it is free on TokenRa until October 31, 2026.

• M3.1 Flash ahead: Automation-Bench v1.0.6, 62.8 against 60.5 for V4.1 Flash-1017. MiniMax published no 1017 score for its other wins, so these rows compare with the earlier V4.1 Flash: BankerToolBench 76.2 vs 73.7, GDP.pdf AA 20.8 vs 12.8, OSWorld 2.0 61.7 vs 54.3, PostTrain Bench 39.3 vs 30.5, SWE Refactor Bench 23.0 vs 11.8 and CharXiv RQ 82.5 vs 69.0.
• DeepSeek V4.1 Flash-1017 ahead: DeepSWE v1.1 77.0 vs 73.7, NL2Repo-Bench 68.2 vs 61.2, Terminal Bench 4.0 42.3 vs 38.1, BabyVision 92.5 vs 91.0 and HLE 44.0 vs 43.4.
• Agents' Last Exam is not like-for-like: MiniMax lists 29.6 for M3.1 Flash and leaves DeepSeek blank, while the DeepSeek-V4.1-Flash-1017 model card reports 35.4 under DeepSeek's own setup.
• How to choose: start with M3.1 Flash for tool-heavy office agents and computer use, start with V4.1 Flash-1017 for repository-scale code generation and terminal agents, and run both on 20 to 50 of your own tasks before you commit.

Where MiniMax M3.1 Flash fits

• Tool-calling agents that need many steps inside a cost budget, where output price dominates total spend.
• Coding assistants that read a repository, propose a patch, and verify it against a test run.
• Computer-use automation that drives interfaces and needs visual grounding plus planning.
• Image and video understanding such as screenshots, charts, and screen recordings, through the same endpoint as text requests.
• High-volume extraction and classification where a Flash-tier model is enough and a frontier model would be wasted.
• Second-opinion routing: send the hard 10% of requests to a larger model and the rest here.

Production checklist

• Keep API keys on the server and separate them by environment.
• Re-check the 1M-token context and enabled input modalities against the live console before depending on them.
• Run vendor-reported benchmarks' neighbouring workloads, not the benchmarks themselves; they are a proxy, not a test.
• Set explicit timeouts, bounded retries, tracing, and an application-side spend cap.
• Given reported hallucination concerns, require source-grounded prompts and verify structured output before acting on it.
• Review provider terms, privacy requirements, and output rights for your use case.

MiniMax M3.1 Flash API FAQ

Continue exploring