Ling 3.2 Flash vs MiniMax M3.1 Flash: The Same 1M Window, Two Different Bills

Put Ling 3.2 Flash and MiniMax M3.1 Flash side by side on a spec sheet and they look like siblings. Both are native multimodal mixture-of-experts models that take text, image and video and return text. Both carry a 1M-token context window. Both activate roughly 23 to 25 billion parameters per token out of a much larger total. Both ship an Anthropic-compatible endpoint alongside an OpenAI-compatible one. The divergence is entirely in what you can find out about them and what you can do with them: MiniMax M3.1 Flash has been callable at a published $0.15 per million input and $0.60 per million output since 19 October 2026, with downloadable weights and an independent intelligence score; Ling 3.2 Flash, live since 5 November, has a free window closing around 19 November, no published rate at all, no weights and no third-party score.

That makes this a comparison of two purchase decisions rather than two models. One you can price, audit, self-host and fail over today. The other you can evaluate free for a few more days and then guess at.

What each one actually is

Ling 3.2 Flash is Ant Ling's flagship multimodal entry: 560B total parameters, 25B active, 1M context, native image and video input, documented on Ant's developer site alongside a text-only sibling at the identical size. It sits at the top of a line whose previous flagship, Ling-3.1-flash, is 560B/25B with the same window and no vision. The 3.2 release is best understood as the moment the top of the Ling line and the multimodal capability landed in the same weights — Ling-3.0-flash-VL had vision, but it was a 124B/5.5B side branch.

MiniMax M3.1 Flash is a different lineage entirely: 428B total, 23B active, 1M context, a native multimodal MoE built on MiniMax Sparse Attention with DSpark speculative decoding and quantization-aware training. It is the successor to MiniMax-M3, which is still in the catalogue at $0.30 input and $1.20 output — exactly double M3.1 Flash on both sides. MiniMax's own documentation describes M3.1 Flash as "an ultra-efficient open-weight model for everyday work," positioned for agentic reasoning, tool use, coding and long-context tasks, with tunable thinking depth across low, medium, high, xhigh and max.

• Parameters — Ling 3.2 Flash 560B total / 25B active vs MiniMax M3.1 Flash 428B / 23B.

• Context — 1M tokens either way; both serve it natively rather than via retrieval.

• Inputs — text, image and video on both; text output on both.

• Max output — MiniMax documents 524,288 tokens; Ant publishes no output ceiling for Ling 3.2 Flash.

• Weights — MiniMax: downloadable, roughly 252 GB, under the MINIMAX COMMUNITY LICENSE. Ling 3.2 Flash: none published.

• Price — MiniMax $0.15 / $0.60 per million with a 90% cache discount vs Ling 3.2 Flash unlisted, free for fourteen days from 5 November.

• Independent score — MiniMax 29 on the Artificial Analysis Intelligence Index v5.0.1 vs Ling 3.2 Flash: no page, no score.

The free window is a deadline, not a discount

Ant's pricing page, last updated the day of launch, has rows for Ling-3.1-flash, Ling-3.0-flash-VL, Ling-3.0-flash and two Ming image models. It has no row for Ling 3.2 Flash. The model appears once, in a promotional line: free for the first two weeks after launch. When that lapses — around 19 November, which is about five days from today — the page says original prices resume and advises watching your balance. That is the whole disclosure.

The neighbouring row is the only estimate available. Ling-3.1-flash lists at ¥2.00 per million input and ¥6.00 output, discounted 75% to ¥0.50 and ¥1.50, cache reads ¥0.40 to ¥0.10. Roughly $0.28/$0.85 at list, $0.07/$0.21 discounted. If Ling 3.2 Flash prices like its text-only sibling it lands near there — but that is inference from an adjacent row, and no Ant page says how image or video tokens are billed. Thirty-two frames from a thirty-second clip are an input Ant has not priced.

MiniMax M3.1 Flash's number, by contrast, is on a page and in our catalogue: $0.15 per million in, $0.60 out, $0.015 per million on cache reads after a 90% discount. That cache line matters more than it looks for agent workloads, because a long system prompt and a stable tool schema re-sent every turn is exactly what caching is for — at $0.015 it is nearly free to keep context parked.

The independent number, and what it costs per task

Artificial Analysis scores MiniMax M3.1 Flash at 29 on the current v5.0.1 Intelligence Index, ranked 7th of 28 models in its comparison set and above the median of 24. The cost to run that index once is $0.49, and the run took 16.3 minutes per task against a median of 14.4 — slower than average. Output speed is 141 tokens per second, and time to first answer token is 53.75 seconds, which is the number that should worry you most: 53.75 against a median of 17.52 means this model thinks before it speaks, and a chat UI will feel it.

It is also verbose. The index run generated 161M output tokens against a median of 106M. Cheap per token, chatty per answer — which is why the per-task cost matters more than the per-token price, and why $0.49 is the figure to carry into a budget rather than $0.15.

Ling 3.2 Flash has none of this. There is no Artificial Analysis page for it as of 14 November; the model URL 404s. The nearest thing in the family is Ling 3.1 Flash 1019 at 29.6 on the same v5.0.1 index, priced at $0.30 input and $0.90 output on the third-party market — a real independent number for the predecessor, and not a number for this model. The only performance evidence for Ling 3.2 Flash is Ant's own launch comparison against Opus 5 at maximum effort: 92.21 against 90.56 on OmniDocBench1.5, 43.71 against 39.52 on ZEROBench(Sub), 85.86 against 80.20 on Vision2Web. Vendor-reported, unreproduced, and all three visual by construction.

So the honest scoreboard has one side with a third-party figure and a cost per task, and the other side with three vendor numbers and no referee. That asymmetry is not a knock on Ling 3.2 Flash's quality — it is nine days old — but it is a fact about what you can decide today.

Two ceilings that will shape your architecture

Here is where the spec sheets stop looking alike. Ant documents its rate limit plainly: a global two requests per second, 429s above that, and a 90-second timeout on synchronous calls with streaming recommended. For an agent that makes dozens of calls per task, two QPS is not a soft limit — it is the throughput number that decides how many users one integration serves.

Ant's multimodal limits are equally specific and equally binding. Images: JPEG or PNG, base64 only, up to forty per request, 32 MiB per encoded string, 32 MiB total request body. Video: MP4, MOV or WMV, thirty seconds maximum, one video per request, fixed 2fps sampling capped at 32 frames, 32 MiB base64 ceiling. There is no URL-fetch path documented for either medium. A "long-horizon multimodal agent" that can hold a visual context across many tool calls is a compelling pitch; a thirty-second, thirty-two-frame, one-video-per-request ceiling is the shape of the loop it can actually run.

MiniMax publishes no equivalent per-request video cap in the documentation we read, and its context story is the simpler one: 1M in, up to 524,288 out, with thinking depth dialled by an effort parameter rather than a separate mode. If your task is "read a long screen recording and produce a long report," the difference in disclosed limits is the difference between a documentation problem and a design problem.

Both models sit behind the same two endpoint dialects, so switching costs are a base URL and a model string — which is precisely the case for not hard-wiring either one. On OrcaRouter, MiniMax M3.1 Flash is one identifier among 200+ on a single key, priced at the provider's own list rate with zero markup, so a MiniMax price change reaches your invoice the day it happens rather than after a repricing cycle. Automatic failover means a 429 or a 90-second timeout on one provider is a retry elsewhere rather than a failed request — the failure mode Ant's own rate-limit page warns about is the one this removes. Ling 3.2 Flash is not on our catalogue, so we will not pretend to serve it; the honest framing is that it competes here for the workloads where MiniMax M3.1 Flash is already one call away.

The licence question nobody puts in the headline

MiniMax M3.1 Flash ships open weights under the MINIMAX COMMUNITY LICENSE, and the terms are worth reading before anyone in the room says "open source." The grant is for non-commercial use. Commercial use requires prominently displaying "Built with MiniMax M3.1-Flash" on a related website, interface, blog post, about page or product documentation, and requires a one-time notice to MiniMax — or prior written authorization if your products generate more than $20 million in yearly revenue. That is a materially different proposition from an MIT licence, and it is the kind of clause that surfaces in legal review two weeks after the prototype works.

Ling 3.2 Flash has no licence to question because it has no weights. The inclusionAI Hugging Face organisation carries Ling-3.1-flash, Ling-3.1-flash-1019 and base checkpoints; as of this week there is no 3.2 repository. Ant's last flagship weight release was 573.8 GB under MIT on 27 October. A lab that open-sources one flagship and closes the next is telling you where it thinks the margin is, and a team that needs to self-host, fine-tune or audit the weights is not served by this release at all.

Which one to pick, and when to pick neither

Pick MiniMax M3.1 Flash if you need a number today. It has a published rate, a 90% cache discount, downloadable weights with a readable licence, an independent score you can compare against 27 other models, and it is callable through us right now. The costs are real: 53.75 seconds to first answer token, 16.3 minutes per index task, and 161M output tokens on a run that a leaner model finishes with two-thirds as many.

Pick Ling 3.2 Flash only if the task is vision-heavy and frame-sized — document understanding with equations and diagrams, drawing-to-3D conversion, render inspection loops — and only for the next five days, while inference is free. It is a 560B native multimodal model with a 1M window and a genuinely unusual benchmark claim on three visual evaluations. What you cannot do is put it behind production: no rate card, no weights, no independent score, two requests per second, and no failover path anywhere including here.

Pick neither if what you actually want is a cheap multimodal model with a stable price and a licence you can live with — the field around these two is deep, and both are mid-tier on the dimensions that decide a production choice: price transparency, throughput ceiling, and the existence of a referee.

The next five days are the whole story for Ling 3.2 Flash. Either Ant publishes a rate card that includes image and video token pricing before the free window closes, and this becomes a real comparison between two priced models, or the window lapses into an unlisted rate and the decision stays one-sided for another quarter. MiniMax M3.1 Flash has already answered every question a buyer asks; Ling 3.2 Flash has answered the interesting one — that a 560B multimodal model can beat Opus 5 on document and 3D benchmarks by its own measurement — and deferred the boring ones, which are the ones that decide whether you can ship.