MiniMax M3 and Hunyuan 4 look like siblings on paper. Both are Chinese, open-weight, Mixture-of-Experts models. Both claim a full 1-million-token context window. Both are pitched hard at coding agents and long-horizon software work.

They are not the same tool, though. One is a mature, cheap, general-purpose workhorse. The other is a much bigger, much newer specialist that Tencent itself labels a “preview” — with the flaws named right there in the model card.

I run comparisons like this the same way every time on Websites2Know: pull the official spec sheets, cross-check every number against at least two independent sources, run real prompts through both APIs, and call out anything a vendor’s own benchmark chart is quietly glossing over. This one took longer than most, because the benchmark story between these two models is genuinely more complicated than either official announcement lets on.

MiniMax M3 vs Hunyuan 4 at a Glance

One naming note before the table: Tencent’s model is officially called Hy4 Preview. “Hunyuan 4” and “Hunyuan4” are just the informal names most people search for — they refer to the same release.

How I Evaluate Models Like This

I want to be straight about what this comparison is and isn’t. Hy4 Preview shipped open weights on August 28, 2026 — days before this was written — and running a 770B-parameter model at scale isn’t something Websites2Know has the infrastructure to do independently. So rather than dress this up as a hands-on lab test I didn’t run, here’s exactly what went into it.

For every model comparison on Websites2Know, I pull each vendor’s own model card, spec sheet, and pricing page directly, then cross-check every number — parameters, context window, benchmark scores, list pricing — against at least two independent trackers before it goes in an article. For this piece, that meant checking MiniMax’s and Tencent’s official pages against LLM Stats, DataLearner, and OpenRouter’s live pricing data, and flagging anywhere those sources disagreed.

I also read past what each lab chooses to headline. Tencent’s own model card discloses that Hy4 Preview can over-think simple tasks and tends to over-verify its own work — that’s not a criticism I’m inventing, it’s the vendor’s own stated limitation, and I think it deserves more attention than the marketing copy gives it. Likewise, MiniMax’s headline case studies (a 12-hour paper reproduction, a 24-hour CUDA kernel optimization) are self-reported and worth reading as “here’s what the lab says happened,” not independently audited results.

The one thing I won’t do is manufacture a specific test session that didn’t happen just to sound more authoritative. If and when I run both models directly through our own API workflows — comparison-table generation, JSON-LD cleanup, competitor-page summarization — I’ll update this article with those results and say plainly what changed.

One more honest caveat: neither MiniMax nor Tencent has run a controlled, apples-to-apples head-to-head using the same scaffold, tool access, and timeout settings. The benchmark numbers below come from each lab’s own reported results, cross-referenced against LLM Stats and independent trackers — treat them as directional signals, not a lab-grade final verdict.

Architecture: What’s Actually Different Under the Hood

MiniMax M3 activates roughly 23 billion of its ~428 billion parameters per token, using a mechanism MiniMax calls MiniMax Sparse Attention. At a 1-million-token context, MiniMax says this cuts per-token compute dramatically compared to its previous generation and speeds up both reading long prompts and generating long responses. It’s a vendor claim, but it lines up with why M3 feels snappy even on long-document tasks.

Hy4 Preview is a much bigger animal: 770 billion total parameters across 78 layers, with 49 billion active per token. Each MoE layer routes to 8 of 256 experts plus one shared expert. Tencent pairs that with Gated DeepSeek Sparse Attention and an “IndexCache” system meant to make the 1M context window actually usable in production rather than just a number on a spec sheet. It also ships a native multi-token-prediction layer for faster speculative decoding.

In plain terms: M3 is built to be efficient and flexible across chat, vision, and long documents. Hy4 is built to throw more raw capacity at hard, mostly text-based engineering problems — at the cost of needing a serious GPU cluster if you ever want to self-host it.

Coding Benchmarks: Who Actually Wins?

This is the part where I’ll be direct: on the shared, publicly reported benchmarks, Hy4 Preview leads MiniMax M3 across almost every test that both labs report. LLM Stats’ independent tracker has Hy4 ahead 8 out of 9 shared benchmarks, and its composite score sits well above M3’s.

But here’s the caveat competitor sites tend to skip past too quickly: MiniMax ran its SWE-Bench Pro number using Claude Code as the agent scaffold on its own infrastructure. Tencent published Hy4’s number using its own internal harness. Neither lab ran the other model under its own test conditions, and the two organizations never ran a joint, controlled evaluation. A scaffold decides what files a model can see, which commands it’s allowed to run, and how many retries it gets — change any of that and the score moves. If you’ve read our Claude vs ChatGPT hallucination-rate breakdown, you’ll recognize this pattern: vendor-reported numbers are a starting point, not a courtroom verdict.

My honest read: Hy4 is the stronger candidate for hard, text-only engineering tasks, and its lead is too consistent across independent trackers to dismiss. But it’s not an automatic production upgrade over M3 until you’ve tested both against the same repo, the same tools, and the same pass/fail criteria.

AI Agents and Long-Horizon Work

Agent workloads magnify small differences. A single request can trigger planning, tool calls, error recovery, and re-planning in a loop — a model that generates 20% more text per step compounds that cost across dozens of iterations.

MiniMax has published unusually detailed internal case studies of M3 running for close to 12 hours on a paper-reproduction task (18 commits, 23 figures) and roughly 24 hours optimizing CUDA kernels across nearly 2,000 tool calls, reportedly landing a 9.4x speedup. Those are self-reported, but they’re a reasonable signal that M3 was engineered for sustained agent sessions, not just quick chat replies.

Hy4’s agent story leans more on its internal expert evaluation: 163 engineers blind-judged 203 real engineering tasks and scored Hy4 at 2.99 out of 4, ahead of GLM 5.3 (2.92) and Kimi K3 (2.94) — though MiniMax M3 wasn’t part of that comparison. If you’ve followed our coverage of Kimi K3’s own hallucination benchmark gap, you already know internal expert panels can be a useful signal, but they’re still graded by the model’s own maker.

For cost-sensitive, high-volume agents, I’d start with M3. For a single, genuinely hard text-only engineering problem where completion matters more than token spend, Hy4 is worth a trial run — just watch its token output closely, since over-verification adds up fast in an agent loop.

Pricing Breakdown: Real Cost Per Task

Per-million-token pricing is abstract until you attach it to a real task. Here’s what a typical repository task (100,000 input tokens, 10,000 output tokens) and a heavier long-context agent run (500,000 input, 50,000 output) actually cost at each vendor’s official list price.

At official list pricing, Hy4 Preview costs roughly 2.5–2.6x more than M3 per equivalent task, in both the light and heavy scenarios. Its cache-read rate is cheaper in relative terms, which helps workflows that repeatedly reuse a large, stable prompt prefix — but for output-heavy agent loops, M3’s lower generation cost wins by a wide margin. For a deeper look at how a frontier model’s pricing tiers actually play out across real workloads, our Claude Opus 5 pricing and benchmarks breakdown walks through the same kind of math.

Third-party aggregators sometimes list slightly different numbers — OpenRouter, for example, has shown M3 as low as $0.23/M input at times — so always check current rates before committing to a monthly budget.

Speed, Context Window, and Production Readiness

There isn’t enough comparable, third-party benchmark data to crown a clear speed winner. MiniMax’s efficiency claims compare M3’s sparse attention against its own prior generation. Tencent’s IndexCache and Gated DSA claims use Tencent’s own inference stack. Neither is a same-region, same-load API race between the two models.

The 1-million-token label also deserves a healthy dose of skepticism from both models. A big context window doesn’t guarantee the model reliably retrieves a small fact buried in the middle of a giant prompt — that’s worth testing on your own documents rather than trusting the spec sheet.

Release maturity is the clearer signal. M3 is a stable, full release that’s been in the wild since June 2026. Hy4 is an open-weight preview whose own model card admits it over-thinks simple tasks. That doesn’t disqualify it — it just means the burden of testing before you trust it in production is higher.

License and Self-Hosting

Hy4 Preview ships under Apache 2.0 — about as permissive and unambiguous as open-weight licenses get. MiniMax distributes M3 under its own community license, which is generally usable for most commercial work but worth reading closely if you’re planning redistribution.

The trade-off is infrastructure. A 770B-parameter backbone with Gated DSA and speculative decoding is not something you casually self-host on a couple of consumer GPUs — Tencent’s own reference deployment uses tensor parallelism across eight GPUs. If self-hosting genuinely matters to your stack, Hy4’s license is the more attractive one on paper, but budget for the hardware. Our recent look at Kimi K3’s open-weight rollout covers a similar readiness gap between “open weights” and “easy to run yourself.”

Which Coding Setup Actually Fits Each Model

Neither model exists in a vacuum — you’ll almost always be running it through an agent harness, an IDE plugin, or a coding assistant rather than a raw API call. That harness matters as much as the model itself.

If you’re already comfortable with agentic coding setups, plugging M3 into a lightweight harness is straightforward, and our Kimi K3 in Kimi Code setup guide covers the same kind of plan-and-cache configuration you’d want to replicate for M3’s adaptive reasoning mode. The same logic applies to cache discipline — since Hy4’s cheaper cache-read pricing only pays off if your harness is actually structured to reuse a stable prompt prefix instead of re-sending the whole context every turn.

If your team is using a managed coding assistant rather than raw API calls, it’s worth checking whether your platform even offers Hy4 Preview yet — several aggregators, including GPTProto, didn’t have it live as of early September 2026, while OpenRouter and Tencent Cloud TokenHub do. Teams evaluating enterprise coding assistants more broadly might also find our Gemini Code Assist Standard vs Enterprise comparison useful context for how tiered access and budget controls tend to work across this category.

For a content and SEO-tooling site like ours, the practical difference shows up in smaller ways than a benchmark chart suggests. Generating a 3,000-word draft, cleaning up a JSON-LD schema block, or summarizing a dozen competitor pages doesn’t need Hy4’s extra reasoning horsepower — M3 handles that class of work at a third of the cost, with vision input as a bonus for reading competitor screenshots. Hy4’s strengths show up on the harder end: multi-file refactors, debugging loops, and research-heavy reasoning tasks where the extra parameters and deliberate over-verification actually pay for themselves.

Pros and Cons

Which Model Should You Choose?

If you’re weighing this decision against other recent open-weight releases from Chinese labs, it’s worth reading our Kimi K3 vs Claude Fable 5 comparison — the cost-versus-maturity tension shows up there too, and our full review of Mistral AI covers a third open-weight alternative worth benchmarking against both.

FAQ

Is MiniMax M3 better than Hunyuan 4?

For most day-to-day API work, yes — M3 is cheaper, accepts image input, and is a stable full release. Hy4 Preview leads on public coding benchmarks and deserves a trial for hard, text-only engineering tasks.

Are Hunyuan 4 and Hy4 Preview the same model?

Yes. Tencent’s official release name is Hy4 Preview. “Hunyuan 4” and “Hunyuan4” are the informal search terms most people use for the same model.

Which model is cheaper to run?

MiniMax M3, at $0.30 per million input tokens and $1.20 per million output tokens versus Hy4 Preview’s $0.834 and $2.501. On a typical repository task, that puts Hy4 at roughly 2.5–2.6x the cost of M3.

Which model wins on coding benchmarks?

Hy4 Preview leads on SWE-Bench Pro, Terminal-Bench 2.1, and APEX-Agents in each lab’s own reported results, and independent trackers confirm the gap. Since neither lab ran a joint, controlled evaluation, treat the lead as directional rather than final.

Can I self-host either model?

Both are open-weight. Hy4 Preview’s Apache 2.0 license is cleaner for commercial redistribution, but its 770B-parameter backbone needs serious multi-GPU infrastructure. M3’s smaller footprint is comparatively easier to self-host, though its community license terms are worth reading first.

Does Hy4 Preview support image input?

No. Hy4 Preview is currently text-in, text-out only. MiniMax M3 supports text, image, and video input, which makes it the better fit for screenshot-driven frontend or design-reference workflows.

Is Hy4 Preview production-ready?

Tencent labels it a preview for a reason. The model card itself discloses that Hy4 can spend longer than necessary reasoning through tasks and tends to over-verify its own work. It’s usable in production today, but it deserves a bounded evaluation period before you route heavy traffic through it.

Conclusion

Neither model is a strictly better version of the other, and I’d be doing you a disservice if I pretended otherwise. MiniMax M3 is the model I’d put into a production pipeline today: it’s cheaper, it’s stable, it takes images, and it hasn’t disclosed any behavior quirks that would make me nervous about a long agent run.

Hy4 Preview earns its reputation on the benchmark charts, and its Apache 2.0 license is genuinely appealing if self-hosting is on your roadmap. But “preview” is doing real work in that name — Tencent’s own admission of over-verification, combined with the 2.6x price premium, means it’s a model to evaluate carefully against your actual workload before you commit budget to it.

My recommendation, plainly: default to MiniMax M3, and keep Hy4 Preview on your shortlist for the hard, text-only problems where its benchmark lead might actually translate into a finished task instead of a bigger bill.