MiniMax M3 vs DeepSeek V3.5: Best open-weights model?
A quick warning before you read on: the original draft of this comparison rests on a model that does not appear to exist. "DeepSeek V3.5" returns no real release. DeepSeek's actual 2026 open-weights lineup is V3.2 and V4 (V4-Pro and V4-Flash, shipped 24 April 2026 under MIT). So treat every "DeepSeek V3.5" figure below as unconfirmed and almost certainly mislabelled. The MiniMax M3 numbers, by contrast, mostly check out.
With that caveat in place: the pitch was that MiniMax M3 and "DeepSeek V3.5" were the two strongest open-weights models you could run in June 2026, both with 1M-token context windows, both openly licensed, and both priced far below the closed models from OpenAI and Google. The real story is narrower. MiniMax M3 is genuine, released 1 June 2026, open-weight, 1M context. Its sparring partner here is not.
For Australian teams weighing an open model to self-host or run cheaply through an API, that distinction matters. You can act on the MiniMax M3 details. The DeepSeek side needs to be re-checked against V4-Pro or V4-Flash before you put a dollar behind it.
Head-to-head benchmarks
| Metric | MiniMax M3 | DeepSeek V3.5 (unverified) | Delta |
|---|---|---|---|
| SWE-bench Pro | 59.0% | 52.4% (unconfirmed) | +6.6 pts (MiniMax) |
| MMLU | 86.4% (unverified) | 85.8% (unconfirmed) | +0.6 pts (MiniMax) |
| Context window | 1M | 1M | , |
| Price (input) | $0.30 / 1M | $0.15 / 1M (unconfirmed) | DeepSeek 2x cheaper |
| Price (output) | $1.20 / 1M | $0.60 / 1M (unconfirmed) | DeepSeek 2x cheaper |
| Licence | Open | Open | , |
One row holds up cleanly. OpenRouter lists MiniMax M3 at $0.30 per 1M input tokens and $1.20 per 1M output, which matches the table. The DeepSeek pricing of $0.15/$0.60 lines up with no real DeepSeek model: V4-Flash sits at roughly $0.14/$0.28, V4-Pro at about $0.435/$0.87, and V3.2 near $0.23/$0.34. So the "2x cheaper" framing rests on a price that does not exist.
Where MiniMax M3 wins
Coding. MiniMax M3's SWE-bench Pro result is the strongest claim in the piece. VentureBeat reports M3 at 59.0% on SWE-bench Pro, narrowly ahead of GPT-5.5 at 58.6%. Worth noting that this is a vendor-run benchmark, so read it as MiniMax's own scorecard rather than an independent audit. The 6.6-point lead over "DeepSeek V3.5" should be ignored, since the comparison model is fictional. If you want a real benchmark fight, line M3 up against DeepSeek V4-Pro, which third-party reviews put around 55.4% on the same test.
General knowledge. The 86.4% MMLU figure for M3 is unverified, no public source confirms it, and most M3 coverage focuses on coding and agentic tasks rather than MMLU. The supposed 0.6-point edge over "DeepSeek V3.5" is meaningless given the other number is attached to a model that does not exist.
Where DeepSeek V3.5 wins
Price. This whole section depends on the unconfirmed $0.15/$0.60 figure, so take it lightly. The original argument was that DeepSeek would be half the price of MiniMax M3, and that the gap compounds on high-volume work like document processing, content analysis, or monitoring. The worked example claimed 100M tokens would cost $15,000 on MiniMax M3 versus $7,500 on DeepSeek. That sum is built on the fabricated DeepSeek pricing and an unusual assumption that bills all 100M tokens at output-style rates, so it does not hold up. If cost is your deciding factor, price it against a real DeepSeek model, V4-Flash in particular is genuinely cheap.
Inference efficiency. The draft claimed DeepSeek was more parameter-efficient and pushed higher throughput on the same hardware "in our testing." There's no published benchmark behind that, and again it points at a model that does not exist, so treat it as unconfirmed. The real DeepSeek V4 does lean on Compressed Sparse Attention for efficiency, but that's a different model and a different claim.
The 1M context parity
The 1M context point is the one part of the comparison that survives, even if the labelling is off. MiniMax M3's 1M-token window is confirmed. DeepSeek's real current model, V4, also ships a 1M-token window, so the parity is genuine, it's just that the matching model is V4, not the "V3.5" named here. The claim that both held accuracy at the far edges of their context windows came from the author's own long-context testing, with no methodology or independent eval attached, so treat that as unconfirmed too.
Verdict
Strip out the fiction and what's left is one model you can actually evaluate. MiniMax M3 is real, openly licensed, runs a confirmed 1M context, lists at $0.30/$1.20 on OpenRouter, and posts a vendor-reported 59.0% on SWE-bench Pro. That's a credible open coding model on its own terms.
The "DeepSeek V3.5" half of this comparison should not drive any decision. If you're shopping DeepSeek, look at V4-Pro, V4-Flash, or V3.2 and pull their real prices and benchmarks before you commit. The headline question, which open-weights model is best, is worth asking, but it needs two models that exist to answer it.