How to Run MiniMax M3.1 Flash via API in 2026 (Every Host Priced)

You can run MiniMax M3.1 Flash through any OpenAI-compatible API by changing two lines — the base URL and the model ID. It is a 428-billion-parameter open-weight mixture-of-experts model with roughly 23 billion active per token, priced at $0.15 per million input tokens and $0.60 per million output tokens on MiniMax's own platform since its October 19, 2026 launch, and served by at least a dozen third-party hosts since the weights landed on Hugging Face on October 23. This guide covers what it is, what it costs on every host type, the license fine print, and how to self-host it if you must.

Key takeaways

• Architecture: ~428B total parameters, ~23B active per token (MiniMax's model card). Sparse MoE, so per-token compute is small — but self-hosting still means fitting all 428B weights in accelerator memory (~252 GB as shipped in MXFP8/NVFP4).
• Context: 1,048,576 tokens with text, image, and video input. Hopper self-hosted deployments cap at 262,144 tokens per the official vLLM recipe.
• Price: $0.15/$0.60 per million tokens list, with cached input at $0.015 — a 90% cache discount. On OpenRouter, Parasail, Fireworks, Together, Nebius Token Factory, DigitalOcean, and MiniMax itself all list the same rate; GMI Cloud lists 25% off.
• License is not MIT: the MiniMax Community License permits commercial use with attribution ("Built with MiniMax M3.1-Flash"), a one-time notice email at or below $20 million yearly revenue, and prior written authorization above it.
• Reasoning is always on: the effort dial runs low → max (default max); there is no none setting, and M3.1 Flash returns its thinking in a separate field rather than M3's <think> tags — an integration change, not a drop-in swap.

What is MiniMax M3.1 Flash?

MiniMax M3.1 Flash is the Flash-tier model in the M3 series from MiniMax, the Shanghai-based AI company that listed on the Hong Kong Stock Exchange in January 2026. It launched on October 19, 2026, positioned for agentic tasks, coding, computer use, and cybersecurity work. An earlier anonymous preview had run on OpenRouter under the name "Space Bunny Alpha", processing more than 200 trillion tokens before MiniMax revealed it as M3.1 Flash.

Three things make it practical for builders. First, the sparse MoE design: cost per token scales with the ~23B active parameters, which is how a 428B model can be priced at Flash-tier rates. Second, MiniMax trained it quantization-aware — dense layers in MXFP8, experts in NVFP4, with a DSpark draft head for speculative decoding — so the released checkpoint is the reference precision, not a shrunken copy of a BF16 original. Third, it is genuinely open-weight: you can download and inspect the files, which matters for teams with data-residency requirements, and it is one of the few frontier-class agentic models where that option exists at all.

The open weights release on October 23 (428B/23B active, on Hugging Face and ModelScope) followed four days after the API launch. Our news coverage of that release digs into the license terms in more detail; this guide focuses on running the model.

How does MiniMax M3.1 Flash benchmark?

Independent numbers first, vendor claims labeled as such.

On Artificial Analysis's Intelligence Index, M3.1 Flash scored 42 at launch (v4.3.2-era scoring) — 13 points above MiniMax M3's 29 and level with GLM-5.3-Flash — at roughly $0.22 per Intelligence Index task, down from about $0.51 for M3. On the current v5.0.1 rebasing it reads 29.1. Artificial Analysis also measured a 93% hallucination rate on AA-Omniscience: the model attempts nearly every question rather than abstaining, which pairs high raw factual accuracy (56%, ahead of DeepSeek V4.1 Flash's 43%) with poor calibrated uncertainty. For open-ended knowledge work without a verification layer, that is the caveat to plan around; for coding and agentic loops with machine-checkable outcomes, it matters less.

MiniMax's own launch materials — vendor-run evaluations, so weight them accordingly — report DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7, CyberGym at 82%, and Automation-Bench at 62.8, and claim nearly 3x faster inference and 70% lower usage costs than M3. Third-party speed measurements vary by host: Artificial Analysis lists MiniMax's own endpoint around 74 tokens per second, while Parasail's tops 148 on OpenRouter's routing.

One weakness MiniMax's table does not mention: M3.1 Flash is a verbose reasoner. Artificial Analysis recorded roughly 180 million output tokens across the index — nearly double the median — which is why its per-token price looks dramatically cheaper than its cost per completed task.

What does the MiniMax M3.1 Flash API cost?

There is no single price: each host sets its own rates. All figures below are dated 2026-11-12, from OpenRouter's provider table and Artificial Analysis; re-check the pricing page before you commit, since this field reprices weekly.

Two cost dynamics matter more than the sticker price:

• Cache economics. The $0.015 cached-input rate is a 90% discount, and agent workloads hit cache hard. OpenRouter's observed effective input price across hosts runs $0.02–$0.07 per million tokens because 85–95% of input tokens cache. If your workflow re-sends long system prompts or repo context, the cache rate — not the base rate — decides your bill.
• Cost per task vs cost per token. The verbose reasoning profile means a task that costs $0.02–$0.05 on M3.1 Flash might cost less in raw tokens on a terser model. MiniMax's own Automation-Bench chart prices M3.1 Flash at roughly $0.045 per task at max effort — a vendor figure — which is excellent for the score, but benchmark your own tasks before assuming the rate card tells the whole story.

No batch tier appeared in MiniMax's launch materials, and pricing is flat across the full 1M context — no long-context surcharge.

Is MiniMax M3.1 Flash's license OK for commercial use?

This is the part most teams skip and shouldn't. "Open weight" is a distribution statement, not a rights grant — our open-weight vs open-source explainer covers the distinction. M3.1 Flash ships under the MiniMax Community License, the same custom license family as M3, and per third-party license tracking it requires:

• Attribution: display the exact phrase "Built with MiniMax M3.1-Flash" prominently on a related website, UI, about page, or documentation. A generic "uses AI" badge does not satisfy it.
• One-time notice at or below $20M: if the products and services using the model generate $20 million or less in yearly revenue, a single notice email to MiniMax keeps you compliant. This applies to small teams too — the notice is not just for big companies.
• Prior written authorization above $20M: past that line, you need separate written permission before deploying.
• Prohibited uses: an appendix bars several categories, including any military purpose.
• Derivatives count: fine-tunes and modified versions are covered by the same commercial-use definition.

For the weights published October 23, read the LICENSE file in the official repository yourself — MiniMax revised license terms between releases before (M2.7 to M3 opened commercial use; M3 to H3 added territorial carve-outs), so terms are a moving target. Note that the hosted API and the weights license are separate documents; API use is governed by MiniMax's platform terms, and third-party hosts apply their own terms on top.

A pre-deployment checklist:

• Read the LICENSE file in the repo revision you actually download, and archive it with your deployment records.
• Decide where the attribution phrase goes and keep evidence of it.
• Document your yearly-revenue calculation for the affected products and send the one-time notice if you're at or below $20M.
• If you serve the model as a hosted API yourself, check whether additional conditions apply — this is where the license is least explicit.
• If you use a third-party host instead of self-hosting, the host's terms govern your data, not the model license.

Where can you run MiniMax M3.1 Flash?

Hosted APIs

• MiniMax first-party (platform.minimax.io) is the reference endpoint: launch-day availability, full 1M context, image/video input, the effort dial, and the M Plan subscription as an alternative to pay-as-you-go.
• OpenRouter routes to 10+ hosts (Parasail, Fireworks, Together, Nebius, DigitalOcean, GMI Cloud, StreamLake, inference.net, SambaNova, MiniMax, and more) through one API, with automatic failover and per-host pricing transparency.
• Specialist hosts — Fireworks, Together, Novita, Nebius, SiliconFlow, SambaNova, Modular, DigitalOcean — all serve it serverless. Our provider pricing cards cover several of these in depth.
• Decentralized options include Morpheus, a decentralized inference marketplace where independent operators serve open-weight models; evaluate it on the same criteria as any host — model availability, published rates, context window, and data policy.

Evaluate every host on identical criteria: does the catalog actually list MiniMax-M3.1-Flash, what context length does the endpoint expose (hosts sometimes cap below the model maximum), what are the rate limits on your tier, where is it served from, and what is the retention policy. Retention and "we don't train on your prompts" statements are self-published policy claims, not independently verified facts. MiniMax's public API documents do not state blanket zero-retention or no-training commitments — third-party privacy reviews have flagged that MiniMax's terms authorize service improvement uses of inputs and outputs — so if privacy is a hard requirement, get any commitment in writing or self-host.

Self-hosting MiniMax M3.1 Flash

The official vLLM recipe branches on GPU generation:

• NVIDIA Blackwell (B200/B300/GB200/GB300): 4 GPUs at tensor-parallel 4, native NVFP4 execution, full 1,048,576-token context, NVFP4 KV cache, DSpark speculative decoding at 7 draft tokens. vLLM's minimal launch: vllm serve MiniMaxAI/MiniMax-M3.1-Flash --tensor-parallel-size 4 --max-model-len 1048576 (dedicated container image vllm/vllm-openai:minimax-m31-flash).
• NVIDIA Hopper (H100/H200): a full 8-GPU node at tensor-parallel 8, context capped at 262,144 tokens, video input disabled — Hopper lacks native FP4 tensor cores, so the experts run through a Marlin path.
• AMD MI355X: 4 GPUs at tensor-parallel 4 with ROCm-specific settings, per the same recipe.

The checkpoint is about 252 GB across 48 safetensors shards plus the DSpark drafter. Budget beyond the weights: KV cache grows with context and concurrency, and the recipe's memory-utilization targets (0.85–0.92) leave headroom for it. vLLM shipped day-0 support with NVFP4 MegaMoE kernels; SGLang and NVIDIA NIM recipes also exist.

Honest math: 4x B200 spot rents around $23/hr as of early November 2026 — roughly $17k/month before networking and ops. At the API's list rates, $17k buys on the order of 28 billion output tokens (far more with cache hits). Self-hosting only wins at sustained, high-utilization volume, for data-residency mandates, or when you need to fine-tune. Community GGUF quantizations appeared within hours of the weights release, but full-precision quality on consumer hardware is not a realistic target for a 428B model.

How do you call the MiniMax M3.1 Flash API?

Every host above exposes an OpenAI-compatible endpoint, so integration is a two-line change. Model IDs vary (minimax-m3.1-flash on OpenRouter, MiniMax-M3.1-Flash on MiniMax's platform) — copy the exact string from your provider's docs.

Python:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_PROVIDER_KEY",
    base_url="https://YOUR-PROVIDER-ENDPOINT/v1",  # the only line that changes
)

resp = client.chat.completions.create(
    model="minimax-m3.1-flash",  # exact model ID from your provider's catalog
    messages=[{"role": "user", "content": "Explain MoE routing in one paragraph."}],
    stream=True,
)
for chunk in resp:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

curl:

curl https://YOUR-PROVIDER-ENDPOINT/v1/chat/completions \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-m3.1-flash",
    "messages": [{"role": "user", "content": "Explain MoE routing in one paragraph."}],
    "stream": true
  }'

Two integration gotchas:

• Reasoning output moved. M3 returned thinking inline in <think> tags; M3.1 Flash returns it in a separate field. If your parser keys off <think>, update it before switching models. Keep reasoning between turns where your host supports it — dropping it degrades quality, per MiniMax's own guidance across the M series.
• The effort dial. Reasoning is always on. Set effort from low to max (default max); sending effort: "none" errors. Lower settings cut output tokens — and therefore cost per task — at some quality cost, per MiniMax's own effort-level charts.

Provider quirks to check before production: tool-calling schema support (most hosts expose it, some don't), whether structured outputs work on your route, rate limits on your tier, and whether video input is enabled on your endpoint.

MiniMax M3.1 Flash vs M3: should you switch?

Mechanically, switching from M3 to M3.1 Flash is a model-ID change on the same base URL — but it is not a drop-in swap, for three reasons:

1. Reasoning parsing changed (separate field, not <think> tags), per the gotcha above.
2. The economics moved in your favor on agentic work: Artificial Analysis measured cost per Intelligence Index task falling from ~$0.51 to ~$0.22 — a ~57% cut at higher scores — and MiniMax claims 70% lower usage costs, a claim consistent in direction with the independent number but not independently verified.
3. The abstention profile didn't improve. The 93% hallucination rate on AA-Omniscience is an M3.1 Flash measurement, not an M3 comparison — but if you rely on the model to say "I don't know", test that behavior explicitly before switching.

M3 remains listed on many hosts at $0.30/$1.20 (together with a $0.06 cache rate on some). If M3 meets your quality bar and your prompts are tuned to it, there is no forced march — but for new agentic work, M3.1 Flash's price-performance is the reason it replaced M3 as MiniMax's current model after M3 launched in June 2026.

FAQ

How much does the MiniMax M3.1 Flash API cost? MiniMax's list price is $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens, with no long-context surcharge across the 1M window. Most OpenRouter hosts match that list; GMI Cloud listed 25% off at the time of writing. Prices change — check each provider's page.

Is MiniMax M3.1 Flash's license OK for commercial use? It ships under the MiniMax Community License, not MIT or Apache 2.0. Commercial use requires displaying "Built with MiniMax M3.1-Flash", a one-time notice email at or below $20 million in yearly revenue, and prior written authorization above that line. Military use is prohibited. Read the LICENSE file before deploying.

What hardware do you need to self-host MiniMax M3.1 Flash? The official vLLM recipe calls for 4 Blackwell GPUs (B200/B300/GB200/GB300) at tensor-parallel 4 with the full 1,048,576-token context, or a full 8-GPU H100/H200 node capped at 262,144 tokens with video input disabled. The checkpoint is about 252 GB. Consumer GPUs are not a realistic path.

What is the difference between MiniMax M3.1 Flash and M3? MiniMax says M3.1 Flash is nearly 3x faster and about 70% cheaper per task than M3, and Artificial Analysis measured cost per Intelligence Index task falling from about $0.51 to about $0.22. M3.1 Flash also ships quantization-aware (MXFP8/NVFP4) where M3 shipped BF16, and moves thinking output to a separate field instead of <think> tags.

What is MiniMax M3.1 Flash's context window? The model supports 1,048,576 tokens with text, image, and video input. Hopper-GPU self-hosted deployments cap context at 262,144 tokens, and some third-party hosts expose shorter windows, so check the endpoint you call, not just the model card.