MiniMax M3 Review: Open-Weight Model That Competes with Claude Opus? (2025)
I've been watching MiniMax's M series for a while, and the M3 just landed with some wild claims. Frontier coding performance, million-token context, native multimodality — all in an open-weight model. Here's my honest take.
·10 min read
MiniMax M3
8.5/10
Open-weight frontier model with coding, agentic, and multimodal capabilities
"A potential game-changer for open-weight AI, if the benchmarks hold up in practice."
✓ Pros
✓Scored 83.5 on BrowseComp, surpassing Opus 4.7 (79.3)
✓1M token context window with guaranteed 512K minimum
✓Autonomously replicated an ICLR 2025 paper in 12 hours
✓Optimized CUDA kernel to 9.4x speedup without human intervention
✓Natively multimodal from training start, not a bolt-on
✓Fully open-source and open-weight on HuggingFace soon
✗ Cons
✗Parameter count unknown, likely too large for local deployment on consumer hardware
✗Limited real-world community validation so far
✗Previous MiniMax models had mixed reception; trust is not automatic
✗API pricing details unclear, may not be competitive for smaller projects
Introduction
I've been burned by overhyped model announcements before. Every few months, some lab drops a paper with benchmark charts that look too good to be true — and half the time, they are.
So when MiniMax announced M3 — claiming frontier coding performance, a million-token context window, native multimodality, and all in an open-weight model — I did what any sensible developer would do. I raised an eyebrow.
But then I looked at the numbers. 83.5 on BrowseComp, beating Claude Opus 4.7. Autonomous ICLR paper replication over 12 hours. A 9.4x CUDA kernel speedup with zero human intervention. Those aren't incremental improvements — those are leaps.
This isn't a full review from hands-on testing (I haven't gotten access yet), but a deep dive into what MiniMax is claiming, how it stacks up, and whether you should care. Spoiler: you probably should.
What Matters When Evaluating AI Models
Before diving into M3, let's set the criteria I use for evaluating any new model:
Coding ability — can it write and debug real code, not just LeetCode? Agentic reasoning — can it break down complex tasks, use tools, and persist over long horizons? Context length — raw max is one thing, but can it actually use that context effectively? Multimodality — is it a first-class capability or a hack? Openness — can I run it myself, fine-tune it, trust it? Cost — both API pricing and hardware requirements.
M3 claims to hit all six. That's rare even among closed-source frontier models.
MiniMax M3 — The Details
M3 is built on MiniMax's proprietary Sparse Attention (MSA) architecture, which is the secret sauce behind the massive context window. The API supports up to 1M tokens with a guaranteed minimum of 512K — that's enough to hold an entire codebase or a long video in one go.
It's natively multimodal. MiniMax rebuilt the entire data pipeline to scale pretraining to 100T+ tokens, with multimodal training from the start. That means the model doesn't just bolt on vision — it understands images, charts, and video as core capabilities.
MiniMax M3 vs. Top Models
Feature
RecommendedMiniMax M3Token plan
Claude Opus 4.7$15/1M input
GPT-5.5$10/1M input
Qwen3-235BFree (self-host)
SWE-bench Verified (coding)
Top-tier (exact % pending)
42.4%
39.3%
~35%
Agentic tasks (τ-bench)
Strong (autonomous replication)
Very strong
Strong
Good
Context window
1M tokens
200K tokens
128K tokens
128K tokens
Multimodal vision
Yes
Yes
Yes
No
Open-source
Yes (soon)
No
No
Yes
API cost ($/1M input tokens)
See site
$15
$10
Free
Get Started
* Links marked with → are affiliate links. We earn a commission at no extra cost to you.
Ad
Advertisement
The benchmarks are where it gets interesting:
—BrowseComp: 83.5 vs. Opus 4.7's 79.3
—Paper Reproduction: M3 autonomously replicated an ICLR 2025 outstanding paper over 12 hours, producing 18 commits and 23 experimental figures.
—CUDA Kernel Optimization: Given only a task description and a non-runnable Triton skeleton, M3 completed 147 benchmark submissions and 1,959 tool calls over ~24 hours, pushing hardware utilization from 7.6% to 71.3% — a 9.4x speedup.
—PostTrainBench: M3 trained four pretrain-only base models on math, code, and knowledge QA, scoring 37.1 — #3 overall, behind only Opus 4.7 and GPT-5.5.
These aren't cherry-picked metrics. They demonstrate genuine coding and agentic capability.
Where It Falls Short
Let's get the caveats out of the way.
First, we still don't know the parameter count. MiniMax hasn't disclosed it. Based on API pricing and comparisons, it's likely much larger than M2.7 — maybe 500B+ parameters. That means local deployment on consumer hardware is probably out of reach. The open-weight announcement is great, but if you need a cluster to run it, it's not exactly accessible.
Second, previous MiniMax models had a mixed reception. The M2.5 and M3 (wait, M2.5?) — actually M1 and M2.5 — were decent but never the community favorites. M2.7 improved, but many found it mediocre at writing and general tasks. Trust has to be rebuilt.
Third, I haven't seen independent third-party benchmarks yet. The numbers come from MiniMax's own evaluations. I want to see results on live code tests, real-world agent tasks, and long-context retrieval before I'm fully convinced.
Finally, the API pricing is opaque. They mention a token plan with a 10% referral discount, but no per-token costs. If it's priced like OpenAI's o1 or Claude Opus, it might not be the budget option you're hoping for.
How It Compares to Claude Opus and GPT-5.5
MiniMax explicitly positions M3 against Claude Opus 4.7 and GPT-5.5. On coding benchmarks, M3 matches or beats both in specific tests. On agentic tasks, the autonomous paper reproduction and CUDA optimization suggest it's at least in the same league.
But there are differences. Claude Opus is battle-tested in production, with a massive ecosystem of tools and integrations. GPT-5.5 has the brand trust and the broadest API support. M3 is unproven outside of MiniMax's controlled demos.
Where M3 has a clear advantage: context length. 1M tokens is double what Claude offers, and the sparse attention architecture might make it more efficient at using that context. Also, open-weight is a huge deal for companies that want to fine-tune or avoid vendor lock-in.
But for day-to-day coding, I'd still trust Claude or GPT more until I see M3 handle real-world messy codebases without hallucinating.
Runner-Ups Worth Knowing
If M3 doesn't deliver, here are other models I'm watching:
Claude Opus 4.7: Still the gold standard for coding and nuanced reasoning. Expensive, but reliable.
GPT-5.5: Faster and cheaper than Opus, slightly behind on hard coding tasks. Great all-rounder.
Qwen3-235B: The best open-weight alternative right now. Strong coding, decent context, and fits on a single H100. But no multimodality and context is only 128K.
DeepSeek-V3: Another open-weight contender, though coding benchmarks lag behind. Good for general tasks.
The Decision Framework
If you need an open-weight model for coding and agentic tasks, can afford the hardware or API costs, and are willing to be an early adopter — M3 is the most promising option right now.
If you need something that just works today, stick with Claude or GPT. They're proven in production.
If you want to experiment locally, Qwen3-235B is a safer bet until M3's parameter count and hardware requirements are clear.
What I Actually Use
Honestly, I haven't switched to M3 yet. I use Claude for serious coding — its artifact system and reliability are unmatched. GPT for quick research and brainstorming. And Qwen3-235B for local experiments where I need to test something offline.
But I have M3 on my radar. If the open-source release includes a reasonably sized model that runs on a few GPUs, I'll be the first to try it. For now, it's a promise — a very exciting one, but still a promise.
Final Take
MiniMax M3 is the most impressive open-weight announcement I've seen in a long time. The combination of coding, agentic reasoning, long context, and native multimodality is genuinely frontier-level.
But I've been around long enough to know that benchmarks aren't reality. The real test will come when the community gets its hands on the weights, runs it on real-world tasks, and pokes holes in the claims.
If M3 delivers even 80% of what MiniMax is showing, it's a landmark moment. If it falls short, it's still pushing the open-weight frontier forward. Either way, MiniMax is worth paying attention to.
I'll update this review once I've actually used M3. Until then, consider this cautious optimism.