MiniMax M3.1 Flash vs DeepSeek V4.1 Flash: Both Open, Only One Is MIT

Until this week, the argument between MiniMax M3.1 Flash and DeepSeek V4.1 Flash was a benchmark-table argument, because that was all either lab had given anyone to argue about. That ended on 23 October 2026, when MiniMax published the M3.1 Flash weights on Hugging Face, four days after the model launched on 19 October. DeepSeek V4.1 Flash had been downloadable since its 10 September release, and its 1017 checkpoint has been on Hugging Face under MIT since 17 October. Both models are now open weights. Only one of them is open source, and the difference is not academic: it decides whether a company can build a product on the checkpoint without asking anyone's permission.

The capability gap is real but narrow. Artificial Analysis puts DeepSeek V4.1 Flash 1017 at 45 on its Intelligence Index and MiniMax M3.1 Flash at 43, on the same v4.3.2 suite. The economic gap runs the other way: $0.26 to complete an index task on DeepSeek against $0.22 on MiniMax. The licence gap does not run either way. It is a categorical difference, and it is the first thing to settle before the benchmark rows matter.

What changed on 23 October

Before the weights landed, the sensible comparison was the one we published on the vendor tables: two labs, overlapping benchmark names, only a couple of rows measured the same way. That comparison is still valid, and it still says DeepSeek is ahead on the shared coding and agent rows. What it could not answer was the procurement question, because MiniMax had promised weights "later" with no date, no licence and no parameter count attached.

All three of those blanks are now filled. MiniMax M3.1 Flash is a 428B-total, 23B-active mixture-of-experts with a 1,000,000-token context, a 524,288-token output ceiling, and a 252.4 GB repository across 48 shards plus a speculative-decoding drafter. The licence is the MiniMax Community Licence. DeepSeek V4.1 Flash 1017 remains a 552B-total model with 16B active parameters during decode, a 1M-token context, a 384,000-token output ceiling, and a 510.3 GB repository under MIT.

Put those two sentences next to each other and the shape of the decision is clear. MiniMax is the smaller download and the easier model to fit on four GPUs. DeepSeek is the larger download, the stronger model on most shared benchmarks, and the one with a licence a legal team can clear in an afternoon.

The licences, side by side

MIT is the simplest open-source licence in common use: use it, modify it, redistribute it, sell it, with no notification and no revenue test. DeepSeek V4.1 Flash ships under MIT, and the Hugging Face repository carries the licence file. If your product embeds the weights, MIT asks only that you preserve the copyright notice.

The MiniMax Community Licence is a commercial contract dressed as an open licence. It is free for non-commercial use. Commercial use carries two conditions. Every commercial deployment must display "Built with MiniMax M3" prominently — on a website, in a user interface, in a blog post, on an about page, or in product documentation. And the revenue tier decides what else is owed.

• At or below $20 million in annual revenue — a one-time notice to MiniMax by email, subject line specified in the licence

• Above $20 million in annual revenue — prior written authorisation from MiniMax before commercial use

• Commercial hosting — explicitly covered by the licence's definition, so an inference business built on the weights is inside the clause, not outside it

• Fine-tuned and post-trained versions — also covered when deployed commercially

• Prohibited uses — illegal content, military purposes, harm to minors, knowingly spreading false information, and discrimination or hate speech against protected groups

The revenue line is the part that changes behaviour. A two-person startup can comply with an email and a footer link. A company at $30 million cannot ship the model without a negotiation, and the negotiation is with a Chinese lab whose terms may change in the next revision. That is a different risk profile from MIT even if the download button looks the same. It also explains why some commentators have started calling releases of this shape "pseudo-open": the weights are open, the permission is not.

252 GB against 510 GB, and why that is not the whole serving story

The download sizes are the first concrete difference a self-hosting team hits. MiniMax's repository is 252.4 GB because the model is natively trained for NVFP4 routed experts and an NVFP4 KV cache; the checkpoint ships already quantised, not as a BF16 file you compress yourself. DeepSeek's 1017 repository is 510.3 GB, roughly twice the bytes for a model with 124 billion more total parameters but fewer active ones.

Those numbers are not like-for-like — the two labs made different precision choices, and a smaller file does not automatically mean a cheaper deployment — but they are what you actually move, store and load. The serving consequence is more useful than the download size. vLLM's day-0 post reports that M3.1 Flash fits 76 concurrent 256K-token sequences on four NVIDIA B300 GPUs, against 28 for MiniMax-M3 with an FP8 KV cache. It reports 418.9 output tokens per second per user on GB300 AgentX at the latency-sensitive end, and 153.8K total tokens per second per GPU at the throughput end.

DeepSeek's architecture points the other way. Its 552B total is larger, but only 16B parameters are active during decode and 8B during prefill, so the compute per generated token is lower than the total suggests. The trade is memory capacity and interconnect: a 510 GB checkpoint points at a multi-node deployment, while MiniMax has published a four-GPU recipe. MiniMax is easier to stand up on a small cluster. DeepSeek is faster per generated token once the cluster exists, and its off-peak rate matches MiniMax's flat rate.

Capability: 45 against 43, and the rows underneath

On the independent index the two-point gap is real but not a chasm. Both models sit in the top handful of open-weight models, both have a million-token context, and both are reasoning models that spend a lot of output tokens before they answer. The difference is in the shape of the capability, not the headline.

• Intelligence Index — 45 for DeepSeek V4.1 Flash 1017, 43 for MiniMax M3.1 Flash, both on Artificial Analysis v4.3.2

• Cost per index task — $0.26 against $0.22, a 15% edge to MiniMax

• Output speed — 206.1 tokens per second against 92.5, a 2.2x edge to DeepSeek

• Time to first token — 1.01 seconds against 1.52 seconds

• Max output — 384,000 tokens against 524,288

• Context — 1,000,000 tokens on both

• Inputs — text and image on DeepSeek; text, image and video on MiniMax

The vendor benchmarks tell a more specific story, and every one of them is vendor-reported. On the two rows both labs measured on a named, comparable harness, DeepSeek is ahead: DeepSWE v1.1 at 77.0% against 73.7%, and Agents' Last Exam at 35.4 against 29.6. On Terminal-Bench 4.0 DeepSeek reports 42.3 in the Claude Code harness against MiniMax's 38.1. MiniMax wins Automation-Bench v1.0.6 at 62.8 against 60.5, and edges Humanity's Last Exam without tools at 43.4 against 42.6. The rest of each table is measured on a different rig, judged by a different model, or published by only one side.

Then there is the number neither launch post leads with. Artificial Analysis's AA-Omniscience evaluation put MiniMax M3.1 Flash at 16 overall with a 93% hallucination rate: when it did not know, it answered rather than abstained almost every time. DeepSeek's card carries its own warning about sparse-selection errors in its compressed attention and state reconstruction at very long contexts. Neither caveat shows up in a vendor benchmark table, and both matter more in an agent that acts than in a chat window that can be closed.

Price: flat against a clock

The rate cards look identical at the headline and diverge the moment you read the hours. MiniMax charges $0.15 per million input tokens and $0.60 per million output tokens, flat, with cached input at $0.015. DeepSeek charges the same $0.15 and $0.60 off-peak, but doubles to $0.30 and $1.20 between 01:00 and 04:00 UTC and again between 06:00 and 10:00 UTC on weekdays. Its cache hit rate is $0.003 off-peak and $0.006 at peak, a 98% discount.

For a workload that runs on US hours, DeepSeek's peak window is mostly out of the way and the two cards are close. For a European team whose CI runs at 07:00 London time, four hours of every weekday are billed at double, and MiniMax's flat card wins without any capability argument. For a cache-heavy agent loop over a stable repository, DeepSeek's $0.003 cache read is five times cheaper than MiniMax's $0.015, which can swamp the per-token difference. The right comparison is a month of your actual traffic, not a blended rate.

Which checkpoint should you download

Download MiniMax M3.1 Flash if you want the smaller artefact, the four-GPU recipe, the longer output ceiling, video input, or a licence you can satisfy with an email and an attribution line. It is the easier model to stand up on hardware you already have, and its cost per completed index task is the lower of the two.

Download DeepSeek V4.1 Flash 1017 if you want MIT, the stronger shared benchmark rows, roughly twice the output speed, or a cache discount deep enough to change the economics of an agent loop. It is the harder deployment and the easier procurement, which is the opposite trade from MiniMax's.

Do not download either if your monthly token bill is small. At $0.60 per million output tokens, a billion tokens a month is $600 through MiniMax's API; the engineering time to keep a 252 GB checkpoint serving reliably costs more than that for most teams. Open weights are a strategic option, not a default cost saving.

Where OrcaRouter fits

OrcaRouter does not host MiniMax M3.1 Flash today, and this article is not a claim that it does. It does route DeepSeek V4.1 Flash 1017, along with the previous MiniMax-M3 generation, on the same key as more than 200 other models. That matters for a comparison like this one: you can put DeepSeek's checkpoint in front of your own traffic without a second contract, keep MiniMax-M3 as the fallback, and let automatic failover handle the route that wobbles. Provider list prices are passed through at 0% markup, so DeepSeek's peak and off-peak rates show up on the bill as the vendor charges them, not as a blended estimate.

If you self-host either checkpoint, the router is a complement rather than a substitute: it gives you a hosted fallback for the hours your cluster is saturated or being upgraded. If you call the API, it is the difference between a comparison you read and a comparison you can run on your own prompts this afternoon.

The short version: MiniMax has closed the capability gap far enough that licence and serving footprint now decide the choice, and DeepSeek still wins the licence. That is a more interesting matchup than the benchmark table was, and a more durable one.

One last distinction worth keeping straight: the 1017 suffix is not decoration. The September DeepSeek V4.1 Flash checkpoint scores 39 on the same index, four points below MiniMax. Every comparison in this article is against the 17 October checkpoint, because that is the one whose weights, licence and benchmark rows are current.