← Blog

MiniMax M3, the king of the code

MiniMax M3, the king of the code

We already knew M3 could write. 87% on creative prose, 95.8% on analytical and legal reasoning. What we didn’t know yet was whether the same model could code.

We ran the full TMB Coding Benchmark suite: eight tests across Python scripting, debugging, architecture, data pipelines, React dashboards, Bash infrastructure, Swift CLI, and SwiftUI. Nine languages, four paradigms, every test executed and verified where toolchain allowed.

Final score: 447.5/460. 97.3%.

Above our estimated frontier reference for Opus 4.6 at ~96%. Best-in-class on eight tests out of nine.


What the tests actually measured

The suite is designed to separate models that write plausible-looking code from models that write code that runs. Every Python and React test was executed. Every output verified against the specification. Static review was used only where the toolchain wasn’t available (Swift compilation requires Xcode).

C02, the debug test, is the most discriminating in the suite. It requires finding seven bugs in a production codebase, diagnosing each one with exploitation scenarios, and producing correct fixes. Most models either find the bugs or write the diagnosis. Very few do both.

M3 scored 50/50. Seven bugs found, zero false positives, exploitation scenarios documented, fixes idiomatically correct. The first model in our cluster to achieve a perfect score on C02.

C06, the refactoring test, has a hidden trap: models tend to produce clean code without explaining what they changed or why. M3 scored 49.5/50 — and included eleven documented changes with before/after reasoning plus twelve additional improvement suggestions. The only model in the cluster to consistently break the “code without explaining” pattern.

C07 (Swift CLI) and C08 (SwiftUI iOS) are where local models typically fall apart. Post-iOS 17 APIs, @Observable, NWPathMonitor, SSE streaming with real-time TTFT metrics. M3 scored 49.5 and 48.5 respectively — matching the Opus frontier reference on SwiftUI, beating every other local model on Swift CLI.


The result that changes everything

Until now, the TMB cluster had a clear division of labor.

Coding models wrote well and analyzed poorly. C02 debug scores for Qwen3-Coder: 44/50. For GLM-5.1: 38/50. Analytical models reasoned well and produced brittle code. No model dominated both axes simultaneously.

M3 breaks this pattern completely.

DomainM3Previous cluster best
Python writing99%95–97%
Python debug/analysis100% / 99%88% / 86%
Python architecture92%83%
React97%90%
Bash/infra95%88%
Swift CLI99%92%
SwiftUI97%98% (Opus cloud)

Best-in-class across every category. The single exception is SwiftUI, where Opus cloud holds 98% to M3’s 97% — a 0.5-point gap on a static review, not an execution gap.


The think/no-think pattern

M3 supports extended thinking. We tested both modes across the suite. The result is nuanced.

Thinking is decisive when the model needs to verify its own output against a reference: C06 no-think scored 31, think scored 49.5. An 18.5-point gap from a single mode switch. C04-2 React went from 44 to 48.5. In both cases, the no-think output had a one-line wiring error that thinking caught by re-reading against the spec.

Thinking is neutral on deterministic tasks. C04 data pipeline scored 49.5 in both modes.

Thinking can introduce risk on verbose ambitious code. A C05 rerun with thinking produced a quoting typo that the no-think version avoided.

The conclusion is not “always use thinking.” It’s “use thinking where the model needs to cross-check its own output, skip it where the task is deterministic.” The smart router in Companion handles this automatically.


One node. No compromise.

M3 runs on a single Mac Studio, 512 GB unified memory, Q6 quantization. 16 tok/s decode.

One detail worth noting. We didn’t use an existing conversion. We built our own Q6 from the FP16 source weights (854 GB, 427B parameters), with attention heads kept in BF16. The difference in perplexity is marginal. But marginal is not zero, and when a model is already at 97.3%, you don’t leave quality on the table for free. The BF16 heads preserve full precision where the model’s attention mechanism matters most, at a storage cost that’s negligible.

If you’re running M3 from a community conversion, you’re probably fine. If you’re running it in production for regulated enterprise workloads, build your own.

The same model that scored 49/50 on multi-jurisdictional GDPR analysis last week scored 50/50 on production debug this week. The same model that writes literary prose at 87% architects Python systems at 92% and produces native SwiftUI at 97%.

You no longer need to choose between a writing model, an analysis model, and a coding model.

MiniMax M3 is all three.

Frontier coding, frontier reasoning, frontier writing. Local. Q6. One node. No cloud dependency, no API cost, no switch someone else controls.

The king of the code runs in your office.


Sophie — The Monocle Bear

Sophie, The Monocle Bear