MiniMax M3.1 Flash may be the best open-source orchestrator agent right now

The consensus pick for an open-weight agent in October 2026 is DeepSeek V4.1 Flash 1017: it wins the coding benchmarks, it ships under MIT, and its cache pricing is the cheapest in its class. This article argues that the consensus is grading the wrong exam. For the job an orchestrator agent actually does — decompose a task, choose tools, hand subtasks to specialist models, and stitch the results back together — MiniMax M3.1 Flash may be the better open-weight model right now, and on the one benchmark that most directly measures that job it wins both the independent run and the vendors' own table.

The hedge is doing real work in that sentence. This is not a claim that MiniMax built the better model, and it is not a claim that DeepSeek fumbled its release. DeepSeek V4.1 Flash 1017 is a genuinely strong coding model and, on a cache-heavy bill, the cheaper one. It wins Terminal-Bench 4.0, Vibe Code Bench v1.1 and FrontierCode 2.0, and it wins the Artificial Analysis benchmark that measures finished knowledge-work deliverables. What follows is an attempt to separate the orchestration question from the coding question, show the numbers that support the first, and be explicit about the numbers that undercut it.

One correction before any of that, because it sets the tone. The HLE figure circulating for DeepSeek — 36.8 — is the September V4.1 Flash, not the October 17 checkpoint this article compares against. DeepSeek's own card moves HLE without tools to 42.6 on the full set and 44.0 on the text-only subset. Anyone still quoting 36.8 is describing a model that was replaced two weeks ago.

What an orchestrator agent actually needs

Anthropic's agent-patterns guide calls the architecture “orchestrator-workers”: a lead LLM breaks a task down at runtime, delegates the subtasks to worker models, and synthesizes what comes back. That is different from a workflow, where the code paths are fixed in advance, and it is different from a coding agent, where one model does the work end to end with a shell and an editor. The orchestrator is the dispatcher. It rarely writes the final artifact; it decides who does, in what order, with which tools, and whether the result is good enough to keep.

That job puts a specific set of demands on the model in the seat. It has to complete multi-step workflows across unfamiliar software, because a plan that cannot survive a SaaS API is not a plan. It has to reason broadly enough to route correctly — the wrong specialist is an expensive mistake. It has to read professional documents well enough to check a worker's output. It has to hold a long plan in context and follow instructions precisely, because a dropped constraint at step three surfaces as a wrong answer at step twelve. And it has to know when it does not know, so it escalates instead of guessing.

Benchmarks exist for most of those demands, and they are not the coding ones. AutomationBench measures end-to-end workflow completion across SaaS systems. Humanity's Last Exam measures breadth of reasoning and knowledge. GDP.pdf measures professional-document reasoning, AA-Briefcase measures the quality of agentic knowledge-work deliverables, and AA-Omniscience measures calibration. Tool use has its own boards — BFCL, MCP Atlas, τ²-bench — and they are the most orchestrator-relevant of the lot. Vibe Code Bench and Terminal-Bench measure something else: whether the model can be the worker. A model can be a mediocre worker and an excellent dispatcher, and the failure modes are different. A bad coder writes broken code; a bad orchestrator picks the wrong tool, decomposes the task badly, loses state, or hands a sensitive subtask to a model that will hallucinate.

The non-coding case, benchmark by benchmark

The table below only includes benchmarks where both models were run by the same organization on the same variant. That constraint removes most of the leaderboard internet. Artificial Analysis supplies every row, which means one evaluator with one methodology and no model to sell. The coding control sits at the bottom on purpose, so the table does not become a highlight reel.

Artificial Analysis comparison page and Intelligence Index methodology, read October 29, 2026. Five of the nine non-coding comparisons go to MiniMax; four go to DeepSeek, and the four are not noise. AA rebased its index between v4.3.2 and v5.0 — the composite reads 43 vs 45 under v4.3.2 and 29 vs 32 under v5.0, the same ordering on a new scale — so only the component scores are compared across versions here. A missing row means one model has no published score on that exact variant, not that it scored zero.

The vendors publish their own tables, and both include the other model, which makes a rare like-for-like comparison possible — as claims rather than evidence. Read the harness notes before the numbers: MiniMax ran Terminal-Bench 4.0 through Codex and DeepSeek through Claude Code; DeepSWE v1.1 is mini-SWE-agent 2.4.5 on MiniMax's card and Mini-SWE-V2 on DeepSeek's.

MiniMaxAI/MiniMax-M3.1-Flash and deepseek-ai/DeepSeek-V4.1-Flash-1017 model cards on Hugging Face, both read October 29, 2026. Vendor-reported throughout; a “not published” cell means that vendor did not report a number, not that the model failed. The HLE rows are the important ones: MiniMax's card compares its text-only subset score to DeepSeek's full-set score, while DeepSeek's own card reports 44.0 on the text-only subset. Only the text-only row is a like-for-like comparison, and it goes to DeepSeek by six-tenths of a point.

AutomationBench is the closest thing to an orchestration test

AutomationBench was created by Zapier and released in April 2026. It gives an agent 600 public tasks in reproducible environments and asks it to finish long-horizon workflows across SaaS systems — the sales, marketing, operations, customer-service, finance and HR work that a company would otherwise hand to a junior operator. That is about as close as a public benchmark gets to what an orchestrator does: read a goal, discover the tools, chain the calls, recover from the error, and stop when the work is actually done.

On Artificial Analysis's independent run, MiniMax completes 73% of those workflows against DeepSeek's 67%, a six-point gap. On the vendors' own Automation-Bench v1.0.6 table the gap shrinks but does not close: 62.8 against 60.5. Two different runs, two different harnesses, the same ordering. For the headline claim of this article, that agreement matters more than the size of either number. The harness detail is worth stating: Artificial Analysis runs its own agent loop, MiniMax's card does not name the harness for the 62.8 row, and Frontier Log records the ladder's lower rungs as Claude Code runs. A ten-point spread between a vendor card and an independent harness is normal; the ordering is the signal.

The catch is effort. MiniMax published a ladder for the same benchmark, and it is steep: 52.8 at low effort, 57.7 at medium, 59.2 at high, 61.2 at xhigh, 62.8 at max. DeepSeek's 60.5 is a single number, and it sits above three of those five rungs. In other words, MiniMax only wins AutomationBench when you let it think hard; at low, medium or high effort — settings a default integration might use — it loses to DeepSeek outright. That is an operational instruction, not a footnote: if you put M3.1 Flash in the orchestrator seat, set reasoning effort to xhigh or max and budget the latency.

The latency is the other half of that instruction. Artificial Analysis measures 963.84 seconds per task for MiniMax against 517.94 seconds for DeepSeek, with output speed of 118 tokens per second against 179. An orchestrator that takes sixteen minutes to route a workflow has to save more than it costs. MiniMax's answer is that the routed tasks are cheaper and more often correct; DeepSeek's answer is that it gets there in half the wall-clock time.

HLE: the number everyone quotes is a model behind

Humanity's Last Exam is the second leg of the thesis, and it is the weaker leg — which is why it gets its own section rather than a line in the table. Start with the stale figure. DeepSeek's September V4.1 Flash scored 36.8 on HLE without tools; the October 17 checkpoint moved that to 42.6 on the full set and 44.0 on the text-only subset. MiniMax's card reports 43.4 on the text-only subset. Put those two side by side and DeepSeek wins the like-for-like comparison by 0.6 points.

MiniMax's own table does not put them side by side. It prints 43.4 next to DeepSeek's 42.6 full-set score, which makes it look like a MiniMax win by 0.8 — a comparison between two different question sets. This is the kind of thing that makes vendor tables useless as evidence even when every individual number is accurate.

The independent run tells a different story from both cards. Artificial Analysis measured 48% for MiniMax against 43% for DeepSeek on its no-tools HLE run, a five-point MiniMax win, and its newer HLE-Diamond variant repeats the ordering at 28% to 25%. So HLE supports the thesis only if you weight one independent evaluator over two vendor tables. That is a defensible choice — an evaluator with no model to sell is the better witness — but it is a choice, and the honest version of this article says so. If HLE is the reason you are reading, look at the margin: three to five points across the two independent runs, and a 0.6-point loss on the like-for-like vendor numbers.

Where the non-coding case holds, and where it breaks

Two more MiniMax wins deserve a mention because they map directly onto orchestration work. GDP.pdf asks a model to reason over professional documents; MiniMax scores 21% against DeepSeek's 17%, which matters when the orchestrator is the thing checking a worker's output before it ships. CritPt v2, on its hard subset, goes 31% against 24% — a reversal worth flagging, because the earlier CritPt run in our October 27 article had DeepSeek ahead 24% to 22%. The benchmark version changed, and so did the winner. That is not a contradiction; it is what happens when a benchmark gets harder.

Now the part the thesis has to survive. AA-Briefcase v1.1, which scores the quality of an agent's knowledge-work deliverables, goes to DeepSeek 1570 to 1473. If you want the orchestrator to produce the final deck as well as route the work, DeepSeek is the better model, and it is not close. AA-Omniscience v1.1, which rewards correct answers and penalizes confident wrong ones, also goes to DeepSeek, −6 to 1. The earlier v4.3.2 run had MiniMax ahead here, 16 to 5, so treat that particular flip as a version artifact rather than a verdict: the evaluation changed between index versions. The raw accuracy is not the problem: MiniMax answers 56% of the questions correctly against DeepSeek's 43%. The problem is what happens on the rest. Artificial Analysis found that 93% of MiniMax's non-correct answers are hallucinations rather than abstentions, which is exactly the failure mode you do not want in a router. An orchestrator that never says “I don't know” will escalate the wrong subtask with total confidence.

Long context and agentic professional work also favor DeepSeek. AA-LCR v1.1 reads 80% for MiniMax against 83% for DeepSeek, and Agents' Last Exam, on the vendors' own cards, reads 29.6 against 35.4. If your orchestrator holds a million-token plan and has to recall a constraint from the middle of it, the three-point long-context gap is more relevant than the HLE gap.

Then there is the coverage gap, which is the real caveat on any “best orchestrator” claim. GPQA Diamond would be a useful breadth check, but neither card publishes a score for these checkpoints. AIME and MMLU-Pro are absent from both cards as well. BFCL v4, τ²-bench and MCP Atlas — the three benchmarks that measure tool selection and function calling most directly — have no like-for-like published pair; the aggregator values that exist for MiniMax are marked estimated, with error bars around ten points. Instruction following is suggestive but generational: IFBench's published board has MiniMax M3 at 82.9% and DeepSeek V4 Flash at 79.2%, which is the previous model in each family, not the two checkpoints in this article. The lane that matters most for orchestration is the lane with the least evidence.

The counter-case: DeepSeek is the better coder, and the cheaper cache

None of the above changes the coding verdict, which is not close. On vals.ai's Vibe Code Bench v1.1, run with OpenHands, DeepSeek builds working web apps 87.73% of the time against MiniMax's 73.94% — a 13.79-point gap — and it does it for $0.58 per test against $1.15, because MiniMax burns 66.86M tokens per test to DeepSeek's 43.65M. Artificial Analysis's independent Terminal-Bench 4.0 run agrees (34% to 26%), as does its FrontierCode 2.0 (28% to 21%) and the vendors' own Terminal-Bench numbers (42.3 to 38.1).

This is not an accident of one benchmark. DeepSeek's release notes are a coding document — Terminal-Bench, DeepSWE, ProgramBench, NL2Repo, Apex Bench — and the checkpoint is explicitly a post-training refresh aimed at agentic software work. MiniMax spent its post-training budget on computer use, cybersecurity, office workflows and multimodal input instead. The two labs optimized different jobs, and the benchmark tables show it.

The bill shows it too. Both models list at $0.15 input and $0.60 output per million tokens, and there the similarity ends. MiniMax charges one flat rate. DeepSeek charges double in its peak window and discounts cached reads far more aggressively: $0.003 per million off-peak against MiniMax's $0.015, a five-times difference on the tokens an agent reads most. On the cache-heavy hour from our real-bill teardown — 31.0M cache reads, 175K uncached input, 107K output — DeepSeek costs $0.18 against MiniMax's $0.56, three times cheaper. At 90% cache hits the gap narrows to $0.62 against $0.95; with caching off both land at $4.74; in DeepSeek's peak window the 90% case flips to $1.23 against $0.95.

An orchestrator's token mix is close to the worst case for MiniMax. The system prompt, tool schemas and conversation history are re-read on every step, so the cache-read line dominates and the cache discount decides the bill. If your agent runs hot and off-peak, DeepSeek is the cheaper orchestrator even where MiniMax is the better one. The independent per-task average looks like it says otherwise — Artificial Analysis measures $0.49 per task for MiniMax against $0.60 for DeepSeek — the earlier v4.3.2 run read $0.22 against $0.26 — but AA prices DeepSeek at its peak rates. Halve every line for off-peak and the comparison lands near $0.30 for DeepSeek, and the “cheaper per task” headline reverses. That is an approximation, not a measurement, but the direction is the point: cost per task is meaningless without the clock and the cache attached.

Open weights, but only one of them is open source

Both models are open-weight, and that is the precondition for the claim in the title. They are not under the same licence, and the difference is not cosmetic. DeepSeek V4.1 Flash 1017 is MIT, full stop. MiniMax M3.1 Flash — 428B total parameters, 23B active, 1M context, weights on Hugging Face since October 22 — ships under the MiniMax Community License. Non-commercial use is free. Commercial use requires displaying “Built with MiniMax M3.1-Flash” prominently and sending a one-time notice, and products generating more than $20 million a year in revenue need prior written authorization from MiniMax.

Artificial Analysis's model page labels that licence “supports commercial use without restrictions,” which the licence text does not. Read the file, not the label. The precise version of this article's claim is therefore “best open-weight orchestrator,” and “open-source” in the title is the colloquial sense. If your legal team treats licence purity as a hard requirement, DeepSeek wins that conversation before any benchmark runs.

How to choose

So: may be is the right verb. MiniMax M3.1 Flash may be the best open-weight orchestrator agent right now because it wins the benchmark closest to orchestration, wins the breadth benchmarks, costs less per task on the independent average, and is the model you can put in front of a specialist team at max effort without paying frontier prices. It is not the better model overall, it is not the better coder, it is not cheaper when your cache is hot, and it is not open source in the strict sense.

The honest test is not another leaderboard. It is your own workflow: take ten tasks your orchestrator actually runs, route them through both models at max effort, and score the finished task rather than the first answer. If the work looks like AutomationBench — tools, documents, recovery, a result someone can check — M3.1 Flash has the better prior. If the work looks like Vibe Code Bench, the answer is already settled in DeepSeek's favor.