DeepSeek V4 Pro vs MiniMax M3: The Open-Weight Flagship Face-Off

DeepSeek V4 Pro and MiniMax M3 are the two open-weight flagships fighting for the coding crown this year, and the interesting thing is how differently they try to win it. MiniMax M3, released June 1, is the first open-weight model to claim frontier coding, a 1-million-token context, and native multimodal input in one package, built on M​iniMax's Sparse Attention architecture. DeepSeek V4 Pro, whose official version landed August 13 after a four-month preview, is the reasoning specialist: a 1.6-trillion-parameter Mixture-of-Experts with 49 billion active parameters, the highest Codeforces rating ever recorded for a language model, and now a much stronger agentic scorecard. Both are MIT-licensed, both are cheap, and both are aggressively positioned against models that cost ten to fifty times as much.

That headline "both open-weight" hides the real divergence. These are not two companies building the same thing with different logos. M3 bets on long-context economics — sparse attention that makes a million-token window cheap enough to actually use. DeepSeek V4 Pro bets on raw reasoning depth — a huge expert-parallel MoE that thinks hard before it answers. Which one is "better for coding" turns out to depend on which part of coding you mean, and on whether your input is a stack trace or a screenshot.

Two architectures, two bets

DeepSeek V4 Pro is the largest open-weights model released to date: 1.6 trillion total parameters with 49 billion active per token, a 61-layer MoE with 384 routed experts (top-6), FP4 expert weights with FP8 attention, and a hybrid attention design that combines compressed and heavily-compressed sparse attention. It needs about 862 GB in FP8, which puts self-hosting in eight-H100 territory rather than on a workstation. The model that is now MIT-licensed and downloadable is the same one behind the official API build that shipped this week.

MiniMax M3 is a smaller and leaner machine — roughly 428 billion total parameters with about 23 billion active per token — built around M​iniMax Sparse Attention (MSA). The vendor claims MSA cuts per-token compute to roughly one-twentieth of the previous generation, with more than 9x faster prefill and more than 15x faster decode at long context. That is the entire point of the model: a million-token window is only useful if inference inside it is affordable, and MSA is what makes that true. M3 also takes image and video input natively, which V4 Pro does not.

Both expose a 1,048,576-token context. DeepSeek V4 Pro tops out at 384K max output; MiniMax M3's ceiling is higher still, at 512K. Both have a toggleable thinking mode — DeepSeek V4 Pro defaults to thinking on, M3's is exposed as a switch.

What DeepSeek's official release changed

Any current comparison that was written before this week is comparing the wrong DeepSeek model. The preview that ran from April 24 was a reasoning model that posted a DeepSWE score of 12.8. The official build, DeepSeek-V4-Pro-0813, pushed that to 62.7 on DeepSeek's own evals, and added a Responses API, Codex integration, and OpenAI- plus Anthropic-compatible endpoints — so the same client code that talks to one can talk to the other. Agentic scores on the vendor's card now sit near Anthropic's Fable 5: Terminal-Bench 2.1 at 87.9 (Fable 5: 88.0), CyberGym 83.3 (Fable 5: 83.1), AutomationBench 31.8 (Fable 5: 29.1). These are DeepSeek-reported numbers on a build that is a day old, unreproduced by any independent lab — treat them as claims, but the claims are no longer modest ones.

The coding scoreboard, and whose numbers are whose

On the most-cited coding benchmarks, the two models split the board, and the split is informative:

• LiveCodeBench — DeepSeek V4 Pro 93.5, the strongest open-weight result on this benchmark and its most-cited win.

• SWE-Bench Verified — DeepSeek V4 Pro 80.6 vs MiniMax M3 80.5, effectively a tie.

• SWE-Bench Pro — MiniMax M3 59.0 vs DeepSeek V4 Pro 55.4, a meaningful 3.6-point M3 lead on agentic software engineering.

• Terminal-Bench 2.1 — DeepSeek V4 Pro 67.9 vs MiniMax M3 66.0, close.

• BrowseComp — MiniMax M3 83.5, a browser-based research agent score that clears many closed models.

• Codeforces — DeepSeek V4 Pro 3206, the highest rating ever recorded for a language model.

The first four of those lines are vendor-reported; the Artificial Analysis Intelligence Index is the one independent number both appear on, and it ranks MiniMax M3 at 44 against DeepSeek V4 Pro at 32 in its non-reasoning configuration (the reasoning config hadn't been scored at AA at the time of writing). A third-party coding bake-off published this month gives the sharper read: MiniMax M3 scored highest overall on a real-world implementation task with a judge praising its "industrial-grade" code, while DeepSeek V4 Pro was described as the best architecture designer but weaker at carrying business logic through to working code — "good at thinking, not at doing." For competitive-programming-style algorithmic problems, the Codeforces and LiveCodeBench numbers say V4 Pro. For turning a spec into a finished, tested service, the real-world results lean M3.

Multimodality is the real fork in the road

Strip away the benchmarks and the deciding difference is input modality. MiniMax M3 accepts text, image, and video natively and supports computer use and visual grounding — an agent can read a diagram, watch a screen, and act on what it sees. DeepSeek V4 Pro is text-only, which is exactly right for pure code and document pipelines and useless for anything where the answer is inside an image. If your coding agent works from issues and repos, V4 Pro's text focus costs you nothing. If it ever needs to look at a UI mockup, a crash screenshot, or a whiteboard photo, M3 is the only one of the two that can.

Price and speed: two different curves

At provider list prices passed through by OrcaRouter, the two models are cheap in different places:

• DeepSeek V4 Pro — $0.44 per million input, $0.88 per million output, cached input $0.0035 per million.

• MiniMax M3 — $0.30 per million input, $1.20 per million output, cached input $0.06 per million.

M3 wins on input and caching; V4 Pro wins on output, and output is where thinking tokens bill. Because DeepSeek V4 Pro defaults to thinking-on and its reasoning is verbose, a long agent run can tip the blended cost toward V4 Pro's side despite the cheaper M3 input. Speed runs the other way: MiniMax M3 measures around 80–145 output tokens per second depending on provider, against DeepSeek V4 Pro's roughly 54–65. M3's sparse-attention advantage shows up most at long context, where its decode stays fast while V4 Pro's attention cost climbs.

Both are on OrcaRouter behind one OpenAI-compatible key, which makes this less of a coin flip than it looks: a production path can route DeepSeek V4 Pro for the algorithmic and long-writing passes, MiniMax M3 for the multimodal and long-context passes, and use the routing DSL or model fusion to have them cooperate in a single call — each billing at the provider list price with 0% markup, with automatic failover if one provider's endpoint wobbles.

Which one fits

• Competitive programming, algorithmic reasoning, math, maximum output per call → DeepSeek V4 Pro. The Codeforces and LiveCodeBench results are the strongest open-weight numbers on the board.

• Turning specs into finished services, long autonomous agent runs → MiniMax M3, per the real-world coding bake-offs and its agentic SWE-Bench Pro lead.

• Any input that is not text (screenshots, diagrams, video stills) → MiniMax M3, the only multimodal option here.

• Million-token context that has to stay affordable and fast → MiniMax M3's sparse attention, at 9x prefill and 15x decode gains over the prior generation.

• Lowest output cost for text → DeepSeek V4 Pro at $0.88 vs M3's $1.20, especially with thinking on.

• Self-hosting on modest datacenter GPUs → MiniMax M3's smaller active footprint; V4 Pro's 862 GB FP8 checkpoint wants a serious cluster.

Neither model has a settled verdict yet, and both are moving fast — DeepSeek V4 Pro's official build landed this week and hasn't been independently benchmarked, MiniMax M3's weights were promised but still need to land fully on the open repos, and DeepSeek has announced a price increase for the V4 line. If you need to pick today, pick on modality and workload: multimodal and long-context agentic work goes to MiniMax M3, pure text-and-code reasoning at the lowest output price goes to DeepSeek V4 Pro — and route them together so that when the first independent scoreboard for either model drops, switching your traffic is a config change, not a rewrite.