MiniMax M3.1 Flash vs Qwen 3.8 Flash: Same Input Price, Opposite Architectures

MiniMax-M3.1-Flash and Qwen3.8-Flash cost the same fifteen cents per million input tokens and are built on incompatible theories of where the money goes. MiniMax-M3.1-Flash, generally available since 19 October 2026, is a closed model that reasons on every call whether you want it to or not and generates at 125.5 tokens per second. Qwen3.8-Flash is a 125-billion-parameter mixture of experts that activates about six billion of them per token, generates at roughly 51.6 tokens per second on the independently measured checkpoint, and sells output tokens for thirteen cents less. One is fast and verbose; the other is slow and terse. Which is cheaper on your traffic depends on a ratio most teams never measure, and that is the only real question in this comparison. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

One thing to fix before the numbers, because it affects every row below. Qwen3.8-Flash is a hosted product. The thing you can download is Qwen3.8-Flash-Next, which shares the architecture but not the name, the licence or necessarily the behaviour — and the independent benchmark figures that exist are for Flash-Next, not for the hosted Flash. MiniMax-M3.1-Flash's only independently measured score is for the model itself, because there is nothing else to measure. Where a figure below is a vendor number it says so; where it is Artificial Analysis, it says so; where the row cannot be compared honestly, the row says that instead.

The two architectures are not two settings of one design

Q​wen's model is the more legible of the two, and its numbers are published in enough detail to reason about. Qwen3.8-Flash is 48 layers with 512 experts per layer, ten routed plus one shared, drawing on 125 billion total parameters with roughly 6 billion active per token — about 95% sparsity. Alongside the expert stack it carries a 51-billion-parameter n-gram embedding table and a 4-billion-parameter multi-token-prediction head, so the stored artefact is closer to 180 billion parameters than to 125 billion. Multi-token prediction is the part that matters for speed: it lets the model propose several tokens per forward pass and verify them, which is how a model this sparse stays usable at all.

M​iniMax has not published a comparable parameter count for MiniMax-M3.1-Flash. The launch post is about outcomes — agentic benchmarks, a claimed 70% reduction in usage cost, inference "nearly three times faster" than the preview — and not about the shape of the network. That asymmetry is itself a data point about the two products: Q​wen is selling an architecture you could in principle host, M​iniMax is selling a service. For a comparison page, it means the architecture argument only runs one way. You can say with confidence why Qwen3.8-Flash is fast per active parameter. You cannot say anything about why MiniMax-M3.1-Flash is fast, because the vendor has not told you.

One line per dimension, both sides on each

• Input price — $0.15 per million tokens on both, which is where the coincidence that frames this page comes from

• Output price — MiniMax-M3.1-Flash at $0.60 per million against Qwen3.8-Flash at $0.47, a 28% gap in M​iniMax's disfavour

• Caching — M​iniMax charges $0.015 per million cached input tokens; QwenCloud lists $0.016 per million cache reads and $0.20 per million for explicit cache creation, so M​iniMax's cache is the simpler and slightly cheaper of the two

• Context — MiniMax-M3.1-Flash at 1,000,000 tokens; Qwen3.8-Flash at 1,000,000 by default with a 991,000-token maximum input and a 131,000-token maximum output

• Independent capability — Artificial Analysis scores MiniMax-M3.1-Flash at 42 on its Intelligence Index, ranked 41st of 173 models; the same tracker scores Qwen3.8-Flash-Next at 40, and no index score is published for the hosted Qwen3.8-Flash

• Output speed — 125.5 tokens per second and a 1.04-second time to first token for MiniMax-M3.1-Flash; 51.6 tokens per second for Qwen3.8-Flash-Next on the same tracker, which is the closest like-for-like figure available

• Cost per completed task — $0.22 for MiniMax-M3.1-Flash on the tracker's own harness; no equivalent figure is published for either Q​wen artefact

• Verbosity — 180 million output tokens consumed across MiniMax-M3.1-Flash's Intelligence Index run, against a 99-million median for models on the tracker; Q​wen publishes no comparable figure

• Reasoning control — M​iniMax returns HTTP 400 if you send thinking: {"type": "disabled"} and rejects reasoning_effort: "none", leaving a five-rung ladder whose default is max; A​libaba documents both thinking and non-thinking modes for the Qwen3.8 family, so this is the one row where the two vendors' documentation is not directly comparable and you should test the toggle on your endpoint rather than trust either page

• Weights — M​iniMax says MiniMax-M3.1-Flash's weights "will be released later" with no date; Qwen3.8-Flash's are not published at all, and the downloadable Qwen3.8-Flash-Next ships under the Q​wen Community License 1.0 rather than Apache 2.0

• Input modalities — text and image on both, with Qwen3.8-Flash also accepting video and both returning text only

The output price is a smaller difference than the output volume

Thirteen cents per million output tokens reads like a rounding error next to a 28% headline. On a real workload it can be the entire decision, in the opposite direction to the rate card, because the two models do not produce the same number of tokens for the same job.

MiniMax-M3.1-Flash cannot stop thinking. Its API reference returns a 400 for the obvious request to disable it, and the only lever is how hard it thinks — low through max, defaulting to max. Artificial Analysis measured the consequence: 180 million output tokens across its index run where the median model on the same tracker spends 99 million. That is 1.8× the output volume of a typical model, and thinking tokens are billed as output tokens and count against your max_tokens ceiling.

Run the arithmetic on a deliberately simple case. A workload that would cost one dollar of output tokens on a median model costs $1.82 on MiniMax-M3.1-Flash if the verbosity ratio holds. The same workload on Qwen3.8-Flash at $0.47 instead of $0.60 costs $0.78 before any volume adjustment — and if Q​wen's non-thinking mode is available and you use it, there is no thinking volume to pay for at all. On output-dominated traffic, the thirteen-cent gap and the 1.8× verbosity multiply into something closer to a 2.3× total difference. That is the case against MiniMax-M3.1-Flash stated as bluntly as it deserves.

The case for it is speed, and speed is not a soft benefit when the workload is agentic. At 125.5 tokens per second against 51.6, MiniMax-M3.1-Flash produces output about 2.4× faster, which means a multi-step tool-use loop that takes Q​wen twenty seconds of generation takes M​iniMax eight. If your agent makes forty model calls to finish a task, that is the difference between a thirteen-minute task and a five-minute one, and latency like that changes what you can build rather than merely how much you pay. It also means M​iniMax's extra tokens arrive in less wall-clock time than Q​wen's fewer tokens, which is the sentence that makes the verbosity argument complicated instead of decisive.

Two vendor benchmark tables you cannot put side by side

Both companies published evaluations, and both sets are unusable for a head-to-head, for a reason worth stating plainly rather than working around.

M​iniMax's table for MiniMax-M3.1-Flash covers BankerToolBench at 76.2, DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7, Agents' Last Exam at 29.6 and CyberGym at 82%, and describes its AutomationBench-AA result as among the top performers without printing a score. A​libaba's table for Qwen3.8-Flash-Next covers SWE-bench Pro at 62.5, DeepSWE 1.1 at 58.7, CoWorkBench at 73.9, AndroidWorld at 84.5, MathVision at 95.7 and JobBench at 55.7. The two lists share exactly one benchmark name — DeepSWE v1.1 — and there the vendor numbers are 73.7% against 58.7%, a fifteen-point gap that looks decisive and is not comparable, because the two figures are for different models than the ones on this page (one is the hosted Flash-Next, the other is M3.1-Flash) and because neither company published the harness, the scaffold or the pass criteria.

What the tables are good for is reading intent. M​iniMax measured computer use, long-horizon professional tasks and vulnerability discovery — the profile of a model meant to operate software. A​libaba measured coding, multi-modal maths and mobile device control — the profile of a model meant to be embedded in applications across modalities. Neither table contains a competitor's number, and neither should be used as though it did.

What the licences actually permit

The open-weight story is the row most buyers get wrong, in both directions.

Qwen3.8-Flash is not open. What is open is Qwen3.8-Flash-Next, and the difference is not cosmetic: it is a separate artefact, published on 26 August 2026, described by A​libaba's own lab as a preview of the architecture that will underpin Qwen4, shipped as 131 shards totalling roughly 360 GB in bf16 and FP8 with no vendor-provided GGUF quantisation. It carries the Q​wen Community License 1.0, which is not Apache 2.0 and is not the licence many teams assume when they hear "Q​wen weights." Anyone planning to download Flash and run it locally should read that licence before sizing the hardware, because 360 GB of weights does not fit on a single accelerator and community quantisations are the only path to a smaller footprint.

MiniMax-M3.1-Flash is closed with a promise. The launch post says the weights "will be released later" and gives no date, no licence and no size. That is a strictly weaker position than Q​wen's — a thing you can download today under a restrictive licence beats a thing you might be able to download eventually under unknown terms — but it is not nothing, and it is the single most likely thing to change the calculus on this page. If M​iniMax ships weights, the price comparison stops mattering to anyone with GPUs, and the sparsity advantage Q​wen bought with 6 billion active parameters faces a competitor that can be served on your own hardware too.

Which one to call today

Both are reachable through OrcaRouter, though not equally. Qwen3.8-Flash is on the catalogue as qwen/qwen3.8-flash, one of more than 200 models behind a single API key with provider list prices passed through at 0% markup. MiniMax-M3.1-Flash is not on the catalogue; it is callable through M​iniMax's own API, its M Plan and M​iniMax Code, and via the serving partner M​iniMax named on 21 October. What is routable from the M​iniMax side is the rest of the line — MiniMax-M3 among them — so a team that wants to test the comparison on its own traffic can put qwen/qwen3.8-flash and minimax/minimax-m3 behind one key and measure, rather than reading two vendor tables and guessing.

That measurement is worth doing before committing, because the two models fail in opposite ways on a benchmark. MiniMax-M3.1-Flash will look expensive per token and cheap per completed task if your work is agentic, with long trajectories that the model finishes in fewer round trips because each one is fast. Qwen3.8-Flash will look cheaper on every line of the invoice and will look worse on any latency-sensitive loop, because 51.6 tokens per second is a real constraint when an agent needs forty sequential replies. Neither of those is a defect. They are the two designs working as intended.

The narrow verdict: if you are building an agent that operates software and you are paying per completed task rather than per token, MiniMax-M3.1-Flash's speed is worth its verbosity, and its closed weights are the price of admission until M​iniMax ships them. If you are running high-volume extraction, classification or generation where output length is bounded and latency is not the constraint, Qwen3.8-Flash is cheaper on the rate card and cheaper again once you stop paying for reasoning tokens you did not ask for — and if you need to own the model rather than rent it, Qwen3.8-Flash-Next is the only one of the pair you can actually hold, under a licence you should read first.