MiniMax M2.7 vs Claude Opus 4.6: 78% SWE-Bench at 10B Active Params

The Headline: 10B Active Params, 78% SWE-Bench Verified

Every coding model release in 2026 has led with SWE-Bench Verified. M2.7's number—78%—is not the absolute leader. Claude Opus 4.6 still sits around 80.84% on the same benchmark, and GPT-5.3-Codex and Gemini 3.1 Pro are clustered within a point. What makes 78% interesting is the rest of the spec sheet:

The right way to read those numbers is not "M2.7 beats Opus." It's "M2.7 gets 95–97% of Opus coding quality for a tiny fraction of the compute." For any team running coding agents at scale, that ratio is the only one that matters at month-end.

For the broader frontier picture, we already covered Claude Opus 4.6 vs GPT-5.3-Codex vs Gemini 3.1 in detail. This post assumes you've seen where the flagships land and are asking the next question: can a 10B-active model realistically replace them as your worker?

Architecture: MoE + Self-Evolving Training on OpenClaw

MiniMax M2.7 is a Mixture-of-Experts model. That part is unremarkable in 2026—most frontier-class open-weight models shipped in the last year have been MoE. What's new is how it was post-trained.

After the base MoE model was pre-trained, MiniMax ran 100+ rounds of autonomous scaffold optimization on the OpenClaw harness. OpenClaw is an open coding-agent scaffold: search, read, patch, test, iterate—the same loop your internal coding agent is almost certainly running in some form. Instead of treating the scaffold as a fixed test harness, MiniMax let the model and the scaffold co-evolve across rounds, using the scaffold's own signals (test failures, patch deltas, tool errors) as training pressure.

This is the idea behind "self-evolving training": you stop optimizing the weights against static coding datasets and start optimizing them against a living agent loop. In practice, the model learns how to use the tools in that loop more than it learns how to write code in isolation. That's why a 10B-active model can close the SWE-Bench gap with 80B-class flagships—it's not better at code in general, it's dramatically better at being the brain inside a specific kind of coding agent.

We've written before about how agent scaffolding beats model upgrades on SWE-Bench—the classic result was a jump from 42% to 78% by improving the harness, not the model. MiniMax M2.7 is the next step: a model co-designed with its scaffold. The upside is cost and latency. The downside is that the benefit is tightest where your agent looks like OpenClaw, and thinner where it doesn't.

Benchmark Head-to-Head

Here's the comparison I'd actually put in front of an engineering lead trying to pick a default coding model today. Numbers for M2.7 are from MiniMax's April 2026 release writeup; numbers for the flagships are the same ones we used in our earlier frontier comparison.

A few things to call out:

What the key data does not give us—and what I will not invent—is a clean public number for MiniMax M2.7 on Terminal-Bench, ARC-AGI-2, or standard reasoning benchmarks. Treat M2.7 as a coding specialist until broader evals exist. If you need a model for general reasoning, architectural planning, or scientific QA, this is not it.

Does Harness-Trained Quality Generalize?

This is the question every engineering lead should ask before standardizing on M2.7, and nobody on launch day has a clean answer yet. Here's my read from shipping agent systems for the last two years.

What transfers well:

What transfers less well:

The mitigation is boring but non-negotiable: run your own evals on your real agent loop, not on SWE-Bench. A 20–50 task internal eval set built from closed tickets tells you more in an afternoon than any public leaderboard. We cover the pattern in evals-driven development; the same logic applies to model swaps.

Decision Framework: M2.7, Opus 4.6, or Gemini 3 Flash

For teams running production coding agents today, this is how I would route work across the three models.

The mental model I give clients: Opus is the architect, M2.7 is the worker, Flash is the intern, and Codex is the ops engineer. Your agent loop should call each of them the same way you'd assign work to people with those roles.

If you're still running everything on a single flagship because routing feels like premature optimization, the cost math has changed enough this quarter that it isn't anymore. We wrote about exactly this in when to use smaller models vs flagship models and again in why specialized 7B models outperform GPT-5 in production. MiniMax M2.7 is the strongest evidence yet that the middle of that routing stack—agent-native, cost-efficient, open-weight-adjacent—is where most real work should live.

What M2.7 Means for the Open-Weight Coding Stack

MiniMax M2.7 lands in the same story arc as DeepSeek V4 and Qwen 3.5: open and semi-open models continuing to close the gap on proprietary flagships not by getting bigger, but by getting more specialized and better-trained. A year ago, the frontier argument was "only the biggest closed models can resolve real GitHub issues." A 10B-active model posting 78% on SWE-Bench Verified retires that argument.

For practitioners, two shifts matter:

At Particula Tech we spend a lot of time building the harness layer on top of models like M2.7 and Opus 4.6—tool design, eval loops, routing, and the unglamorous plumbing that turns a benchmark number into a resolved Jira ticket. If your coding agent is still running everything on a single flagship, there's a real chance the right upgrade this month isn't a new model at all. It's a routed stack with M2.7 doing the work and Opus doing the thinking.

The Bottom Line

MiniMax M2.7 is the first coding model where the "small active parameter count" story is not a compromise pitch. 78% SWE-Bench Verified, 56.22% SWE-Pro, 76.5 SWE Multilingual, and 3x the throughput of Claude Opus 4.6 is a serious production profile—especially for agentic loops that live inside a harness that looks anything like OpenClaw.

Don't replace Opus 4.6 for it. Route to it. Let M2.7 close the 70–80% of tickets that don't need frontier-level reasoning, keep Opus for the ones that do, and drop Gemini 3 Flash underneath both for the trivial edits. Then run your own evals, because the only benchmark that actually decides this is the one built from your own closed tickets.