MiniMax M3.1 Flash Parameters: 428 Billion Total, 23 Billion Active, and the Number That Decides Your Bill

MiniMax M3.1 Flash

What MiniMax actually published

MiniMax released MiniMax M3.1 Flash on 19 October 2026 and put the weights on Hugging Face on 23 October. The model card answers the specification questions the launch post left vague, and several of its numbers differ from the summaries that circulated in between.

• Parameters — 428 billion in total, 23 billion activated per token, in a native multimodal mixture-of-experts architecture.

• Download — 252.4 GB across 48 safetensors shards, plus a 2.3 GB DSpark drafter for speculative decoding.

• Context and output — 1,000,000 input tokens in; up to 524,288 output tokens.

• Modalities — text, image and video in; text out.

• Precision — routed experts and the sparse-attention KV path are trained for NVFP4, shared experts are held in MXFP8, and the query path stays FP8.

• Serving — vLLM, SGLang and TokenSpeed are the recommended paths, with day-0 vLLM containers for NVIDIA and AMD.

• Price on the hosted API — $0.15 per million input tokens, $0.60 output, $0.015 for a cached input read.

• Licence — minimax-community, filed on Hugging Face as code>license: other/code>, which is not MIT and not Apache.

Total versus active: what the 23 billion is doing

A dense 428-billion-parameter model would run every parameter for every token. A mixture-of-experts routes each token to a subset of experts, so the compute per token is set by the 23 billion active figure while the memory footprint is set by the 428 billion total. That split is why the two numbers are usually quoted together, and why quoting only one of them misleads in a predictable direction: the total tells you what hardware you need, the active count tells you roughly what a token costs in compute.

The practical consequence is that 23 billion active is not "a 23-billion-parameter model". It is a model with 428 billion parameters of stored knowledge and 23 billion parameters of per-token work — which is a reasonable description of why it can be both fast to start and comparatively heavy to host.

The 252.4 GB figure is what that looks like in a procurement document: not a theoretical parameter count multiplied by two bytes, but the artefact you move through a security review. The NVFP4 training choice is what keeps it there. In vLLM's accounting the NVFP4 checkpoint is 250.1 GB against 451.5 GB for MiniMax-M3 in MXFP8 — a 44.6% reduction — and the KV cache falls from 61,440 bytes per token to 34,560, a 43.75% cut, because both the sparse-attention indexer and the attention path run in NVFP4.

What the parameters cost at serving time

The number a parameter count cannot give you is concurrency, and that is the one that decides whether self-hosting is viable. On four NVIDIA B300 GPUs at 256,000-token sequences, vLLM reports 76 concurrent sequences for MiniMax M3.1 Flash where MiniMax-M3 with an FP8 KV cache fits 28 — about 2.7 times more. At the latency-sensitive end of its GB300 AgentX run the stack sustains 418.9 output tokens per second per user while serving 38.6K total tokens per second per GPU; at the throughput end it reaches 153.8K total tokens per second per GPU at roughly 127 tokens per second per user. On Rubin the same recipe reports 724.9 tokens per second per user, and on AMD MI355X, 331.

Those are framework benchmarks on specific hardware, published by the framework, not independent production measurements. They are still more decision-relevant than any benchmark row, because a 428B model that fits 76 concurrent 256K sequences on four GPUs is a different procurement object from one that fits eight.

The cost of the 23 billion is not the price of a token

Here is where the parameter story and the invoice diverge. MiniMax M3.1 Flash is priced at $0.15 and $0.60, which is half the input price and roughly a quarter of the output price of Gemini 3.5 Flash-Lite. On Artificial Analysis's cost-per-completed-task measure it still loses: $0.22 against $0.19.

The reason is visible in the same dataset. Across the Intelligence Index the model consumed roughly 180 million output tokens, against a median near 100 million for the models Artificial Analysis tracks. Thinking cannot be disabled — MiniMax's API rejects code>thinking: {"type": "disabled"}/code> with a 400, and the only lever is a five-rung effort ladder from low to max, with max as the default when the field is omitted. Thinking tokens are billed as output and count against code>max_tokens/code>.

So the accurate summary of the parameter economics is this: the active count is small enough to keep the model fast, the verbosity is high enough to keep the invoice from being as small as the rate card suggests, and the two facts are unrelated to each other. On agentic and coding work the deliberation generally earns its tokens. On extraction or classification it does not.

The intelligence result is respectable rather than frontier: 43 on the Artificial Analysis Intelligence Index, fifth of 173 models, with Terminal-Bench 4.0 at 26%, AutomationBench-AA at 73%, GDPval-AA at 1,458 and SciCode at 53%. The abstention result is the one to design around — AA-Omniscience scores it at 16, with a 93% hallucination rate, meaning that when it did not know an answer it answered anyway almost every time rather than declining.

Speed, measured two ways

The published speed figures disagree, and the disagreement is informative rather than alarming. Artificial Analysis measured 1.48 seconds to first token at 95.6 output tokens per second; MiniMax's own material claims a median above 170 tokens per second with peaks above 200. Serving path and measurement conditions move this number a lot, which is why our own route is worth quoting separately: a p50 of 1.29 seconds to first token and a p95 of 4.40 seconds over the last seven days of traffic on code>minimax/minimax-m3.1-flash/code>.

That is the shape a routed model has. The honest way to compare it against an incumbent is to measure your own route rather than trust a single figure from either vendor.

The licence line is part of the specification

Two obligations travel with the weights, and neither appears in a parameter count. Commercial deployments must display the string "Built with MiniMax M3" prominently on a related website, interface, blog post, about page or product documentation — the model is M3.1 Flash and the required attribution still says M3, which is the sort of detail a legal review catches late. And at or below $20 million in annual revenue a company owes MiniMax a one-time notice by email; above that line, prior written authorisation is required before commercial use.

The licence defines commercial use broadly enough to cover a product built on the model, a commercial API on top of it, and a fine-tuned deployment. For a startup under the threshold the cost is an email and a line of attribution. For a larger company it is a licensing conversation on a timeline MiniMax controls. DeepSeek V4.1 Flash ships under MIT and needs no notice at all, which is why "open weights" is not a single category even when both checkpoints are downloadable.

Call it or download it

The break-even on open weights is a utilisation question. At $0.60 per million output tokens, a billion output tokens a month costs $600 through the hosted API; four B300 GPUs cost more than that before power, networking and the engineer who keeps the serving stack current. Downloading wins for data that cannot leave a controlled environment, for fine-tuning runs that need weight access — vLLM reports end-to-end RL with veRL, FSDP2 and rank-32 LoRA on the NVFP4 checkpoint, and a 200-step GRPO run that lifted AIME-2024 accuracy from 36.3% to 67.5% on 16 GB300s — and for batch jobs large enough that a fixed cluster beats a metered bill.

For everyone else the hosted route is the sane default, and MiniMax M3.1 Flash is in the OrcaRouter catalogue as code>minimax/minimax-m3.1-flash/code> at the provider's list price with 0% markup: $0.15 in, $0.60 out, $0.015 cached, one key beside MiniMax-M3 and the models it is usually shortlisted against. Automatic failover covers the case where a provider has a bad afternoon, and the routing DSL can send the hard step of a pipeline to a stronger model while the bulk stays here.

If the parameter question brought you here, the short answer is 428 billion total and 23 billion active. The long answer is that neither number appears on your invoice, and the two that do are the output tokens the model chooses to spend and the licence threshold you may one day cross.

Sourcing note: parameter counts, the 252.4 GB repository, precision scheme, licence terms and vendor benchmark rows are MiniMax's model card and launch material, read on 26 October 2026. The NVFP4 checkpoint sizes, KV-cache bytes per token, concurrency figures, throughput numbers and the GRPO result are vLLM's day-0 post and are framework benchmarks on named hardware. The Intelligence Index score, cost per task, token-consumption estimate and AA-Omniscience figures are Artificial Analysis's. Prices, context and the 1.29-second p50 are OrcaRouter's catalogue and our own seven-day traffic.