MiniMax M3.1 Flash vs Ling 3.1 Flash: One Is Open Source, the Other Is Merely Open Weights

Both of these checkpoints are downloadable. Only one of them is free to build a business on. Ling 3.1 Flash ships under MIT — the licence your legal team approves without a meeting. MiniMax-M3.1-Flash ships under the MiniMax Community License, which is free for non-commercial use, requires a prominent "Built with MiniMax M3" attribution in commercial products, asks for a one-time notice below $20 million of annual revenue, and requires MiniMax's prior written authorisation above it. That clause is the most consequential difference between two models that are otherwise close enough on measurable quality that the gap is inside a rounding error, and it is the difference almost no comparison page puts first.

Everything else about the pair is genuinely close. Ling 3.1 Flash, from InclusionAI and announced on 30 September 2026 with MIT weights following in mid-October, is a 560-billion-parameter hybrid-attention MoE activating about 25 billion per token. MiniMax-M3.1-Flash, from MiniMax and released on 19 October 2026 with weights following on 23 October, is a 428-billion-parameter MoE activating 23 billion per token. Two vendors, two architectures, two weeks apart, and active-parameter counts within eight per cent of each other. The interesting question is not which one is bigger. It is which one you are allowed to use.

Two checkpoints, three weeks apart

Ling 3.1 Flash is Ant Group's InclusionAI lab's model, ID inclusionAI/Ling-3.1-flash, text in and text out. It uses 512 routed experts with 8 selected per token plus a shared expert, and alternates 63 KDA linear-attention layers with 9 gated MLA layers in a 7:1 ratio — an architecture built to keep long-context inference cheap rather than to add modalities. The card claims a one-million-token context; the shipped configuration's max_position_embeddings is 262,144, so treat the larger figure as a stated target rather than a delivered feature. Pricing is $0.30 per million input tokens, $0.90 output and $0.06 on cached reads. Vendor benchmarks exist — Terminal-Bench 4.0 at 40.4%, FrontierSWE at 75.2%, CyberGym at 87.9% — but they were run on Ant's own LingCode harness, which is not public, and none has been independently reproduced.

MiniMax-M3.1-Flash is MiniMax's native multimodal MoE, ID MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face, taking text, image and video and returning text. It activates 23 billion of 428 billion parameters, with sparse attention applied across the whole network, routed experts trained at W4A4 in NVFP4, and a DSpark speculative-decoding drafter bundled in the repository — which is why the download is 252.4 GB rather than the 450-plus a BF16 master would need. Context is 1,000,000 tokens in, 524,288 out. Pricing is $0.15 input, $0.60 output, $0.015 cached. Its vendor benchmarks are similar in kind and equally unreproduced: DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7%, CyberGym at 82.2%, Automation-Bench at 62.8%.

The licence is the fork in the road, not the benchmark

Put the two licence texts next to each other and the comparison stops being technical. MIT means you can fork it, fine-tune it, host it commercially, redistribute the derivative, and never speak to InclusionAI again. You do not need to publish attribution, you do not need to report revenue, and there is no threshold above which someone can revoke your permission. For a company building a product on top of a model, that is the difference between an engineering decision and a procurement cycle.

The MiniMax Community License is not hostile — it is more permissive than a closed API by a wide margin, and most small teams will never touch the $20 million line. But it is a licence with conditions attached to your company's size, and size is the one variable a growing company cannot hold constant. A product that clears the threshold has to go back and negotiate, and the terms of that negotiation are not published. If you are choosing a base model for something you intend to scale, you are choosing a licence that changes meaning at a revenue milestone.

That single fact reorders the rest of this comparison. On capability and price, MiniMax-M3.1-Flash is the stronger proposition. On the question of whether you can build on it without a legal review at the worst possible moment, Ling 3.1 Flash is the one that does not come with a milestone.

The scoreboard, with its caveat attached

Artificial Analysis is the only neutral evaluator that has run both, and its numbers are close enough that the ranking is fragile.

• Intelligence Index — MiniMax-M3.1-Flash 43 against Ling 3.1 Flash 41. A two-point gap, on a tracker where a single eval revision can move a model by that much.

• Cost per index task — $0.22 against $1.00. A 4.5× gap in MiniMax's favour, and wider than the 2× gap on the rate card.

• Input price — $0.15 against $0.30 per million.

• Output price — $0.60 against $0.90 per million.

• Cached input — $0.015 against $0.06 per million. A 4× spread on the line that dominates a cache-heavy agent run.

• Output speed — 95.6 tokens per second against 220.8. This is Ling's clearest win, and it is more than 2×.

• Context window — 1,000,000 tokens against 1,000,000 claimed, 262,144 in the shipped config.

• Maximum output — 524,288 tokens against roughly 33,000.

• Input modalities — text, image and video against text only.

• Licence — MiniMax Community License against MIT.

Why the per-task gap is bigger than the per-token gap

Ling 3.1 Flash is twice the input price and 1.5 times the output price. It is 4.5 times the cost per finished task, and verbosity does not explain it: Ling is the chattier of the two on the tracker, generating 220 million output tokens across the index against 180 million for MiniMax-M3.1-Flash. What explains it is cache pricing and the shape of a reasoning run.

These are both agents that think before answering, which means most of the tokens in a long run are input tokens re-read from cache rather than fresh prompt. On a cache-dominated bill, the line that matters is $0.015 against $0.06, not $0.15 against $0.30. A four-times spread on the dominant cost line turns into a four-and-a-half-times spread on the invoice, and the published rate card understates it by half.

This is also the argument for checking the bill rather than the card. OrcaRouter passes provider list prices through at 0% markup, so a rate change on either side is visible on our side the day it lands rather than after a repricing cycle — and neither Ling 3.1 Flash nor MiniMax-M3.1-Flash is on our catalogue today, so the figures above are the vendors' own published rates and nothing we can route for you. The models we do carry from this tier are MiniMax-M3 and DeepSeek V4.1 Flash, which is what makes the price-per-task experiment cheap to run against a model you can actually call.

Multimodal against text-only is not a feature row

It is a workload filter. MiniMax-M3.1-Flash accepts images and video frames; Ling 3.1 Flash accepts text and nothing else. If your pipeline reads screenshots, parses scanned documents, inspects charts, or looks at video, the comparison ends at the input stage and no amount of MIT licensing compensates for a model that cannot see the input.

Output length cuts the same way. A 524,288-token ceiling against roughly 33,000 is a difference of sixteen times, and it decides any task whose deliverable is itself long — a whole-file rewrite, a repository-scale migration, a structured extraction over a substantial document. A 33K cap does not make those tasks impossible; it makes them a chunking problem you have to solve in your own code, and chunking is where long-context quality quietly degrades.

Where Ling 3.1 Flash is the right pick

Three situations, and they are not niche. First, anything shipping commercially above the $20 million line: MIT removes a negotiation from your roadmap. Second, anything where throughput is the binding constraint — at 220.8 tokens per second against 95.6, Ling finishes a text task in less than half the wall-clock time, and for an interactive agent that is the difference between usable and not. Third, anything where you intend to fine-tune and redistribute: MIT lets you ship the derivative, and the MiniMax licence's attribution requirement follows the output into your product's interface.

Ling is also the cleaner long-context story on paper, in the sense that its hybrid linear attention is designed for it — but the shipped 262,144-token configuration means the one-million-token claim is a promise about a future config, not a property of the checkpoint you download today. Do not budget against the card's number.

Where MiniMax-M3.1-Flash is the right pick

Everywhere the input is not text, and everywhere the output is long. It is also the cheaper model by a margin that survives the licence caveat: at a fifth of Ling's per-task cost, a team under the $20 million threshold can run twice the workload for less than half the money, and the attribution requirement is a line in a footer rather than an engineering constraint. If your product is small, the licence is a formality; if it succeeds, it becomes a conversation. Plan for the conversation rather than pretending it will not happen.

And it is the model with the higher independent score. Two points on the Intelligence Index is not a decisive gap, but it points the same way as the price, the modality support and the context window, which is a pattern rather than a coincidence.

Choosing without choosing twice

The shape this comparison actually implies is not a permanent decision. Ling 3.1 Flash is the better text engine and the safer licence; MiniMax-M3.1-Flash is the better multimodal engine and the cheaper one. A pipeline that needs both — text summarisation at volume, plus document and screenshot handling for the subset that needs it — wants two models behind one interface rather than one model asked to be good at both.

That is the case for a gateway rather than a second contract: one key, one billing surface, automatic failover when either vendor degrades, and the routing DSL when the cost ceiling belongs in the route rather than in application code. Neither of these two models is on OrcaRouter's catalogue, so this is an argument about the shape of the integration, not a claim about what we host. What we do host from the same tier — MiniMax-M3, DeepSeek V4.1 Flash — is there to benchmark the pattern before you commit a production path to it.

The question to answer before either

Not "which scores higher." Two points is noise, and both models will move as evaluators revise their suites. The question is what your company is allowed to build on, and how that answer changes at a revenue milestone you have not reached yet. If the honest answer is that a licence renegotiation would be tolerable, MiniMax-M3.1-Flash is the stronger model at a materially lower price and the decision is easy. If it would not be tolerable, Ling 3.1 Flash gives up two index points, some context headroom and image input in exchange for a licence that never asks you a question — and that is a trade most teams will make once, correctly, rather than discover later.