MiniMax M3.1 Flash Opens Its Weights—and Keeps a Hand on Commercial Use

# MiniMax M3.1 Flash Opens Its Weights—and Keeps a Hand on Commercial Use

*Sophia Chen — October 25, 2026*

MiniMax published the weights for M3.1 Flash on October 23, 2026 — four days after the model's hosted API debut — making a 428-billion-parameter mixture-of-experts checkpoint available for download on Hugging Face↗. The release comes with day-zero support from vLLM and SGLang, a broad hardware partner list spanning NVIDIA, AMD, and domestic Chinese accelerators, and a novel licensing structure that has already sparked debate across the developer community.

That access is meaningful but not unrestricted. The datacenter-oriented model carries a MiniMax Community License requiring attribution, notification, and — above a $20 million annual revenue threshold — prior written authorization from MiniMax before any commercial use. Its technical distribution is broad; its commercial governance remains firmly controlled by the lab.

Four Days That Changed the Nature of the Launch

The October 19 release made M3.1 Flash available through the MiniMax API, M Plan subscription, and MiniMax Code desktop application. Official launch material at that point said weights would follow. On October 23, the Hugging Face repository publication↗ supplied the more definitive event: a downloadable 252.4GB checkpoint under the MiniMax Community License.

Some contemporaneous coverage blurs October 19 and October 23 into a single release date. The distinction matters operationally. October 19 was the hosted and product launch; October 23 was the weights release. For four days, MiniMax's service remained the model's only deployment path. The sequence separated two propositions often treated as equivalent: calling a model via API and possessing, inspecting, modifying, and serving its weights. The latter arrived with different costs and legal conditions.

The earlier "Space Bunny" reveal belongs in this story mainly as a validation and marketing device. MiniMax identified the previously anonymous model — which had been running on OpenRouter since late September, processing over 200 trillion tokens before its identity was disclosed — as an early M3.1 Flash build when it launched the hosted product. Anonymous public use gave the company evidence of demand and real-world operation, but it did not provide reproducible knowledge of the final checkpoint, serving configuration, or benchmark methodology. Publishing weights improves the prospects for independent testing; it does not mean the reported results have already been independently reproduced.

"The Space Bunny stealth period was a calculated move — MiniMax captured 200 trillion tokens of real-world usage data before exposing the model to the volatility of community self-hosting." — Developer analysis, OrcaRouter, October 2026

Architecture: What 428B Parameters Actually Means for Operators

MiniMax describes M3.1 Flash as a native multimodal mixture-of-experts model with 428 billion total parameters and 23 billion activated per token. It accepts text, image, and video input, produces text output, and supports a context window of up to one million tokens.

The 23B active figure is primarily a compute-routing statement. A mixture-of-experts system selects only part of the model for each token, so arithmetic per token can be closer to that active subset than to a dense 428B model. It does not mean operators can download or keep only 23B parameters. The full expert pool must remain available for routing, making total checkpoint size and runtime memory separate from active compute.

The Hugging Face repository↗ is roughly 252.4GB, including 48 weight shards and a DSpark speculative-decoding drafter. That file size is already beyond ordinary consumer-device assumptions, and serving adds memory for the runtime, cache, inputs, and concurrency. A one-million-token maximum context is therefore a capability ceiling, not a promise that million-token requests are inexpensive or practical at high concurrency.

Key Technical Specifications

• Total parameters: 428 billion (MoE architecture); 23 billion active per token
• Context window: Up to 1,000,000 tokens; output limit of 524,288 tokens
• Modalities: Text, image, and video input; text output
• Precision: NVFP4 for routed experts and KV cache; MXFP8 for other paths
• Speculative decoding: Built-in DSpark draft head; community-trained DFlash2 alternative also available
• Checkpoint size: ~252.4GB across 48 safetensors shards
• Reasoning effort: Always-on thinking with five configurable levels (`low`, `medium`, `high`, `xhigh`, `max`); defaults to `max`

MiniMax Sparse Attention replaces the first three full-attention layers from the M3 architecture with sparse-attention paths, reducing bandwidth and decoding pressure at long context. DSpark addresses a different bottleneck: its draft component proposes tokens that the main model can verify, potentially allowing multiple accepted tokens per decoding step. Actual benefit depends on acceptance behavior, batching, and the serving implementation; the presence of a drafter is not itself a guaranteed throughput figure.

Day-Zero Ecosystem Support: Broad Hardware, High Bar

The release arrived with vLLM↗ and SGLang↗ support rather than leaving operators to implement MiniMax Sparse Attention, quantization, and speculative decoding from scratch. AMD's technical documentation↗ confirms support for MI350-series hardware, including MI355X, using ROCm-based AITER kernels with Sequence Parallel prefill and Dual-Batch Overlap optimizations.

This is important for infrastructure teams. Support in established inference engines reduces custom engineering, exposes familiar scheduling and caching controls, and makes evaluation more realistic than a bare checkpoint would. The model is not tied exclusively to NVIDIA, although its NVFP4-oriented design requires hardware-specific treatment where native support differs.

Minimum Deployment Configurations (Published Recipes)

• NVIDIA Hopper: 8× H100 or H200 GPUs (standard throughput configuration)
• NVIDIA Blackwell: 4× B200 or B300 GPUs (NVFP4 native, higher concurrency)
• AMD Instinct: 4× MI355X GPUs (ROCm AITER kernels, Sequence Parallel prefill)
• Recommended frameworks: vLLM, SGLang, TokenSpeed; Baidu FlagOS also supported

The published recipes remain datacenter starting points. Operators must still manage sharding, interconnects, caches, monitoring, and utilization. Day-zero support narrows the integration gap without removing the capital, power, and staffing requirements of a roughly 252GB mixture-of-experts checkpoint. Community quantized versions from Unsloth (GGUF format) lower the hardware floor somewhat, but at the cost of precision and performance that the vendor benchmarks do not reflect.

"Day-zero support from vLLM and SGLang is a generational leap in how Chinese labs coordinate model releases with serving stacks — but it doesn't change the fact that you still need a datacenter to run this thing." — Developer commentary, dev.to, October 2026

The MiniMax Community License: The Decisive Constraint

The most consequential part of this release may be the license rather than the architecture. M3.1 Flash is open-weight, but it is not an unrestricted commercial release and should not be described as permissively open source.

For commercial deployment, the MiniMax Community License↗ requires prominent attribution using the exact wording "Built with MiniMax M3.1-Flash". A company or product with annual revenue at or below $20 million must provide MiniMax with a one-time notification email. Above $20 million, prior written authorization from MiniMax is required before commercial use begins.

License Terms at a Glance

• Attribution required: "Built with MiniMax M3.1-Flash" must appear prominently in product UI, documentation, or website
• Under $20M annual revenue: Commercial use permitted with one-time notification to MiniMax
• Over $20M annual revenue: Prior written authorization from MiniMax required before any commercial use
• Fine-tuned derivatives: Subject to the same commercial-use framework; proprietary post-training does not escape the revenue gate
• Prohibited uses: Military applications, exploitation of minors, generation of harmful false information
• Comparison: Apache 2.0 and MIT licenses impose none of these commercial conditions

The $20 million threshold creates a sharp commercial cliff. A startup can begin below the line with notification and attribution, then require affirmative vendor permission as revenue grows beyond it. That transition may affect fundraising diligence, enterprise procurement, and long-term architecture decisions well before the threshold is crossed. The branding requirement creates another layer of review: consumer applications may be willing to display model attribution, but white-label software, embedded enterprise tools, and products with strict brand controls may not be.

Fine-tuned models do not automatically escape these concerns. The license places modified or fine-tuned derivatives within the commercial-use framework, so proprietary post-training may remain dependent on MiniMax authorization, particularly above the revenue threshold. Developer reaction has been polarized: technical admiration for the architecture and ecosystem support, combined with legal and practical skepticism about the revenue cliff and attribution mandate.

Benchmarks: What the Numbers Actually Say

MiniMax reports 61.7 on OSWorld 2.0 and 73.7% on DeepSWE v1.1. These are vendor figures. OSWorld is directed at computer-use performance, while DeepSWE evaluates software-engineering behavior. Neither score establishes general factual reliability.

Artificial Analysis↗ provides a separate perspective. Its evaluation assigns M3.1 Flash an Intelligence Index score of 42 — up from 29 for the previous M3 generation — while its Omniscience test reports a 93% hallucination rate. The latter concerns behavior on factual questions when the model may not know the answer; it should not be inferred from, or cancelled out by, a coding or computer-operation score. As Shattered.io's analysis↗ notes, when M3.1 Flash provides an incorrect answer, it is "confidently wrong" 93% of the time rather than acknowledging uncertainty.

Benchmark Comparison: M3.1 Flash vs. DeepSeek V4.1 Flash-1017

| Measure | M3.1 Flash | DeepSeek V4.1 Flash-1017 | Source | Signal | |---|---:|---:|---|---| | DeepSWE v1.1 | 73.7% | 77.0% | Vendor-reported | Software engineering | | OSWorld 2.0 | 61.7 | — | Vendor-reported | Computer use | | AA Intelligence Index | 42 | 45 | Independent | Composite capability | | AA Omniscience hallucination | 93% | — | Independent | Factual restraint | | API input price | $0.15/M | $0.14/M | Official | Cost |

The table is not a unified ranking. Prompting, model configuration, reasoning effort, and benchmark versions can differ, and missing cells are not losses. The defensible conclusion is narrower: MiniMax reports competitive agentic and coding performance, while the independent factual evaluation identifies a material reliability risk for open-ended knowledge retrieval tasks.

DeepSeek Offers a Cleaner License Alternative

DeepSeek V4.1 Flash-1017↗ is the most relevant comparison in the Chinese flash tier because it targets similar high-volume coding and agentic work while offering downloadable weights under the MIT License. DeepSeek reports 77.0% on DeepSWE v1.1 and 60.5 on Automation-Bench.

Those figures should not support a categorical rank against M3.1 Flash. The reported configurations are not necessarily identical, and MiniMax's required comparison set does not supply a directly matched Automation-Bench figure. The more robust distinction is licensing: MIT permits commercial use, modification, and redistribution without MiniMax-style revenue authorization or product attribution.

That difference can outweigh a modest benchmark gap. Companies expecting to exceed $20 million in revenue, distribute derivatives, or avoid mandatory vendor branding may find DeepSeek's license structurally easier to approve, even if their own workload evaluation favors MiniMax on some tasks. The VentureBeat analysis↗ captures the tension well: M3.1 Flash beats DeepSeek on certain business benchmarks, but the hallucination rate and license terms complicate the calculus for production deployments.

Self-Hosting vs. API: When Each Makes Sense

• API is likely better when: Volume is moderate or uneven; data residency is not a hard requirement; the team lacks a specialized inference stack; the $20M revenue threshold is not yet a concern
• Self-hosting becomes credible when: Data residency or operational control is mandatory; volume is consistently high enough to justify multi-GPU infrastructure; custom modifications create sufficient value; the organization already operates a datacenter inference stack
• License gate applies in both cases: Above $20M annual revenue, prior written authorization from MiniMax is required regardless of whether you use the API or self-host the weights

The Export Approval Dimension

Bloomberg reported on October 23↗ that both DeepSeek and MiniMax have completed the necessary registration processes under China's interim AI export-control framework — effective October 16, 2026 — and received approval to continue open-sourcing, licensing, and deploying their models overseas. This is regulatory context, not a blanket clearance for every jurisdiction or use case.

The interim framework, issued by MOFCOM and MOST as Announcement No. 48, targets 40 technologies including large language models exceeding two trillion parameters or ranking in the global top ten. M3.1 Flash's 428B total parameters fall below the two-trillion threshold, and the public-domain exemption covers weights already released via open-source publication. Customers still need their own sanctions, export-control, and procurement reviews; the approval does not substitute for those.

Open Distribution, Retained Control

M3.1 Flash gives developers more than an API endpoint: inspectable weights, alternative serving frameworks, and self-hosting capability. Yet its scale confines serious deployment to multi-GPU environments, and its license makes commercial autonomy conditional on revenue level and MiniMax's ongoing authorization.

That is the larger signal for Chinese open-weights releases in 2026. Distribution can accelerate evaluation, ecosystem support, and adoption while the developer retains control over branding, high-revenue users, and downstream commercialization. The MiniMax Community License is not an anomaly — it is a template that other Chinese labs may follow as they seek to balance global developer reach with commercial leverage. M3.1 Flash is open to download, but not open enough to remove MiniMax from many businesses' decisions.

For developers evaluating the model today: the weights are real, the ecosystem support is genuine, and the performance on agentic coding tasks is competitive. The hallucination rate on factual queries is a real constraint for unmonitored deployments. And the license terms require legal review before any commercial use — regardless of whether you plan to self-host or simply call the API.