MiniMax M3.1 Flash Review: 428B Open Weights at $0.15
Quick verdict
MiniMax M3.1 Flash is a 428B-parameter mixture-of-experts model (23B active) with a 1M-token context, text, image and video input, and flat API pricing of $0.15 input / $0.60 output / $0.015 cached input per million tokens — half of M3. It launched October 19, 2026, with weights published October 23 under the MiniMax Community License, which is not an OSI open-source license. Vendor claims include roughly 3x faster output than M3 and 70% lower agent-task cost; Artificial Analysis measures about $0.22 per Intelligence Index task at a score of 43.
What Shipped
MiniMax released M3.1 Flash on October 19, 2026, and published the weights four days later, on October 23. It is the follow-up to M3, aimed squarely at agentic coding, computer use and long tool loops — the workloads where token bills compound fastest.
The model is a native multimodal mixture-of-experts system: about 428 billion total parameters with 23 billion active per token, a 1M-token context, and text, image and video input. MiniMax says the earlier anonymous "Space Bunny" model that topped OpenRouter usage charts in late September was an early version of M3.1 Flash.
Three architecture changes separate it from M3: sparse attention in the first three full-attention layers, quantization-aware training for routed experts and attention paths, and a DSpark speculative-decoding draft head. MiniMax attributes its speed and cost claims to those changes.
Specs at a Glance
| Spec | MiniMax M3.1 Flash |
|---|---|
| Release | October 19, 2026 (weights October 23) |
| Architecture | Native multimodal MoE, ~428B total / 23B active |
| Context window | 1,000,000 tokens |
| Input / output | Text, image, video in; text out |
| Reasoning | Always on; effort low → max, default max |
| API price | $0.15 input / $0.60 output / $0.015 cached input per 1M |
| Model ID | MiniMax-M3.1-Flash (OpenRouter: minimax/minimax-m3.1-flash) |
| Weights | Hugging Face and ModelScope, MiniMax Community License |
| Endpoints | Anthropic-compatible and OpenAI-compatible |
The endpoint compatibility is a practical detail: existing Anthropic or OpenAI SDK code can switch by changing the base URL and model name, which lowers migration cost for teams already running agent loops on either SDK.
Pricing: Half of M3, Flat at Any Length
| Per 1M tokens | M3.1 Flash | M3 (≤512K) | M3 (>512K) |
|---|---|---|---|
| Input | $0.15 | $0.30 | $0.60 |
| Output | $0.60 | $1.20 | $2.40 |
| Cached input | $0.015 | $0.06 | $0.12 |
Pricing is flat regardless of prompt length, which is unusual at a 1M-token window and matters for long-context agents. Against M3, every line is half price below 512K and a quarter above it.
The cost-per-task picture is where the model earns its reputation. Artificial Analysis measures roughly $0.22 per Intelligence Index task against about $0.51 for M3 — a 57% drop — at a score of 43. That puts M3.1 Flash in the same cost band as Claude Haiku 5.5 at max effort ($0.21) and DeepSeek V4.1 Flash at max ($0.26), while offering a much larger context window and multimodal input.
Benchmarks: Vendor Numbers, One Independent Cost Figure
MiniMax's published results, as reported in Chinese cloud-partner coverage, are strong for the price:
- OSWorld V2 (computer use): 61.7 — MiniMax describes it as the highest domestic score.
- BankerToolBench: 76.2 — a banking-workflow tool-use benchmark.
- CyberGym: 82% — security-focused agent tasks.
- DeepSWE v1.1: 73.7 — software engineering tasks.
Treat all four as vendor runs until independent reproductions appear. The one independent figure worth anchoring on is the Artificial Analysis cost-per-task number above, because it measures what a finished task costs rather than what a token costs.
MiniMax also claims more than 100 trillion tokens processed in the first 72 hours after launch, and output speeds above 200 tokens per second at peak with a median above 170. Independent measurements have generally landed below the vendor's speed claims, so size capacity plans against your own throughput tests rather than the headline.
The License Is the Catch
"Open weights" and "open source" are not the same thing here. The MiniMax Community License is non-commercial by default, adds conditions for commercial deployment, and requires written authorization once annual revenue passes $20 million.
For a research lab, a hobby deployment or an internal tool, the terms are workable. For a product company, the license is a business decision, not a footnote — budget legal review before you build a service on the weights. If permissive licensing is a hard requirement, DeepSeek's MIT-licensed models remain the safer choice.
Speed and Self-Hosting
The weights ship already quantized — NVFP4 for routed experts, MXFP8 for dense layers and attention — at roughly 252GB across 48 shards, plus a 2.3GB speculative-decoding drafter. That is small enough for a 256GB unified-memory workstation to hold and large enough that most teams will still prefer the API.
Self-hosting M3.1 Flash makes sense in three cases: data cannot leave your infrastructure, you need to fine-tune, or your volume is high enough that API margins become the dominant cost. Otherwise the $0.15/$0.60 API is cheaper than the hardware, the power and the operations time.
Access and Migration
The model is MiniMax-M3.1-Flash on MiniMax's platform, minimax/minimax-m3.1-flash on OpenRouter, and it is the text model behind all three M Plan subscription tiers (Go, Explore, Build) and MiniMax Code.
Two compatible endpoints lower migration cost: an Anthropic-compatible base URL (api.minimax.io/anthropic) and an OpenAI-compatible one (api.minimax.io/v1). Existing SDK code can switch by changing the base URL and model name.
One integration detail differs from OpenAI-style APIs: because reasoning is always on, the full assistant message — including the thinking block or reasoning_content — has to be passed back in the next request. Strip it and the chain of reasoning breaks between tool calls. Budget output tokens accordingly, since thinking is billed as output.
Limitations
- License. Non-commercial by default; commercial terms and a $20M revenue clause.
- Always-on reasoning. Simple calls pay for thinking whether they need it or not.
- Vendor-run benchmarks. No independent reproduction of the OSWorld or CyberGym scores yet.
- Speed claims. Median throughput in independent tests has trailed MiniMax's figures.
- Hardware floor. About 252GB of weights before serving overhead.
Who Should Use MiniMax M3.1 Flash
- High-volume agent builders running long tool loops where cost per task dominates.
- Multimodal batch pipelines — screenshot analysis, video-frame triage, document review — that benefit from 1M context at a flat rate.
- Teams with data-sovereignty constraints willing to accept the community license.
- Not for frontier reasoning, work that requires a permissive license, or teams that cannot tolerate always-on thinking latency.
Final Verdict
MiniMax M3.1 Flash is the cheapest credible agent-grade model with downloadable weights right now, and the price is not a temporary promotion: it is half of M3 on every line, flat across a 1M-token window.
The trade is explicit. You accept a non-commercial-default license, always-on reasoning and vendor-reported benchmarks in exchange for a cost per task that undercuts nearly everything in its score band. For batch and agent workloads, that is a good trade. For a product you intend to commercialize at scale, read the license first — it is the one part of this release that is not cheap.
Related Articles
Frequently Asked Questions
How much does MiniMax M3.1 Flash cost?+
Is MiniMax M3.1 Flash open source?+
How does M3.1 Flash compare with MiniMax M3?+
What context window and modalities does M3.1 Flash support?+
Does M3.1 Flash always reason?+
What is MiniMax M3.1 Flash best for?+
Pros
- $0.15 input / $0.60 output / $0.015 cached input per million — half of M3 and among the cheapest agent-grade models
- 428B total / 23B active MoE with a 1M-token context and text, image and video input
- Downloadable weights since October 23 for teams that need self-hosting
- Five reasoning-effort levels with Anthropic- and OpenAI-compatible API endpoints
- Vendor claims roughly 3x faster output than M3 and 70% lower cost on agent tasks
Cons
- The MiniMax Community License is not open source: non-commercial by default, with commercial conditions including written authorization above $20M revenue
- Reasoning cannot be disabled, which raises latency and token use on simple calls
- Vendor speed and benchmark claims have not been independently matched
- Pre-quantized weights need roughly 252GB of storage before serving costs are considered
- Independent evidence is limited mostly to cost-per-task figures
Keep reading