MiniMax open-sourced its Code CLI around September 18, 2026 — and the release comes with a genuinely strong benchmark result attached: running on Kimi K3, the tool scored 76.7% (23 of 30 tasks) on the FrontierHarness Eval benchmark, the highest recorded pass rate on that benchmark to date, at $1.83 per pass and the fastest median solve time among tools tested. For builders comparing open-source coding-agent CLIs, this is a genuinely new, verified data point worth adding to the list.
TL;DR
| Question | Answer |
|---|---|
| What was released? | MiniMax's Code CLI, open-sourced |
| When? | Approximately September 18, 2026 |
| Benchmark | FrontierHarness Eval |
| Score | 76.7% (23/30 tasks) — highest recorded on this benchmark |
| Model used | Kimi K3 (from Moonshot AI, run as the backing model for this eval) |
| Cost per pass | $1.83 |
| Speed | Fastest median solve time recorded — 4m33s |
What was actually open-sourced
MiniMax's Code CLI is a command-line coding-agent tool, and open-sourcing it means the tooling — the harness that structures how tasks are broken down, executed, and verified — is now publicly available on GitHub rather than remaining a closed, proprietary product. That's a meaningfully different thing than open-sourcing a model: the underlying model this specific evaluation ran against, Kimi K3, comes from a separate lab (Moonshot AI), not MiniMax itself. What MiniMax is contributing to the open ecosystem here is the CLI and harness layer — the part of a coding agent that determines how effectively a given model's raw capability gets translated into successfully completed coding tasks.
Why the benchmark result is worth taking seriously
A 76.7% pass rate being the highest recorded on FrontierHarness Eval is a real, specific claim worth putting in context: it means this combination of MiniMax's CLI harness and Kimi K3 as the backing model outperformed whatever other tool-and-model combinations had previously been run against the same 30-task benchmark. The additional data points — $1.83 per passing task and the fastest median solve time (4 minutes 33 seconds) — make the result more credible than a pass-rate number alone would be, since a tool that's merely slower or more expensive could technically still claim a similar pass rate through brute-force retrying; this result reportedly leads on all three dimensions simultaneously (accuracy, cost, and speed) rather than trading one off against the others.
What FrontierHarness Eval measures, and why it matters that it's open
FrontierHarness Eval is itself an open-source benchmark project, tracked via GitHub pull requests rather than a closed, institutionally-gated leaderboard. That openness matters for the same reason it matters for any benchmark claim in this space: anyone can inspect the exact 30 tasks, the scoring methodology, and the submission history directly, rather than taking a vendor's self-reported number on faith. For builders evaluating coding-agent tools, an openly inspectable benchmark with a transparent submission history is a meaningfully more trustworthy signal than a closed, vendor-controlled benchmark claim — worth checking the actual GitHub PR directly if you're making a real tooling decision based on this result.
Why open-sourcing the harness matters more than open-sourcing a model
It's worth being specific about why open-sourcing a coding-agent CLI is a genuinely different kind of contribution than open-sourcing a model checkpoint, even though model releases tend to get more attention. A model's raw capability is only one input into how well a coding agent actually performs on real tasks — the harness layer around it (how tasks get decomposed into steps, how the agent verifies its own intermediate progress, how it decides when to retry versus give up, how context gets managed across a long multi-file task) often matters just as much to real-world success rate as the underlying model's raw benchmark score. That's exactly the thesis behind the "harness engineering" framing explainx.ai has covered as its own emerging discipline — the idea that the scaffolding around a model is now a legitimate, separately-optimizable engineering problem, not just a thin wrapper. MiniMax open-sourcing its harness means other teams and open-source contributors can now inspect, fork, and improve that specific layer directly, independent of which underlying model they choose to run it against — a meaningfully more reusable contribution to the ecosystem than a single model release, since it can in principle be paired with future models as they're released rather than being tied to one specific checkpoint's capabilities.
What to actually test if you're evaluating it
For a team seriously considering MiniMax's Code CLI, the practical next step is running it against your own representative coding tasks rather than taking the FrontierHarness Eval number as sufficient on its own — 30 benchmark tasks, however well-designed, can't fully represent the diversity of a real production codebase's quirks, existing conventions, and edge cases. Worth specifically checking: how the harness handles a codebase with an unfamiliar or unconventional project structure, how gracefully it degrades or asks for clarification when a task is genuinely ambiguous rather than confidently guessing, and whether the reported cost-per-pass figure holds up on tasks meaningfully larger or more complex than the benchmark's own 30-task set. Those are the questions a leaderboard number alone can't answer, and they're the ones that actually determine whether a tool like this earns a permanent slot in your team's workflow.
Why this release fits a broader Chinese open-source coding-tool wave
MiniMax's Code CLI release lands amid a broader wave of Chinese AI labs open-sourcing coding-agent tooling and models throughout 2026 — Zhipu's GLM series, MiniMax's own earlier model releases, and multiple other labs have followed a similar playbook of pairing an open release with a specific, verifiable benchmark claim rather than relying on marketing alone to establish credibility. That pattern is worth understanding as a deliberate competitive strategy, not a coincidence: in a market where Western labs like Anthropic and OpenAI have historically commanded more default trust among Western developers, an open-source release with a transparent, independently-inspectable benchmark result is one of the more effective ways for a Chinese lab to establish credibility quickly, since developers can verify the specific claim themselves rather than needing to extend trust based on brand reputation alone. MiniMax's specific combination here — highest pass rate, lowest cost, fastest speed, all on an openly auditable benchmark — is a particularly strong version of that strategy, and it's part of why this release is worth taking seriously as a genuine tooling option rather than dismissing it as one more incremental open-source coding-CLI release in an already crowded field.
Honest limitations
- MiniMax's own official announcement post, beyond the GitHub PR and secondary press coverage, wasn't located at time of writing — verify directly against MiniMax's own channels for the canonical release details.
- This is one benchmark's result, on 30 tasks — a real, useful signal, but not a comprehensive measure of coding-agent capability across the much broader range of real-world software engineering tasks a team might actually need.
- The evaluation is a GitHub PR submission, not an independently-run third-party audit — the standard caveat for any benchmark result applies: verify the methodology and task set directly before treating the specific percentage as definitive for your own use case.
- This post doesn't independently verify the $1.83/pass and 4m33s figures beyond citing the same source — check the linked GitHub PR and evaluation project directly for the full cost and timing methodology.
- How MiniMax's Code CLI performs with backing models other than Kimi K3 isn't addressed by this specific benchmark result — a team pairing the harness with a different underlying model should expect the specific pass rate, cost, and speed figures to shift, since all three are jointly a function of both the harness and the model driving it.
What this means for builders
If you're evaluating open-source coding-agent CLIs, MiniMax's Code CLI paired with Kimi K3 is now a concrete, benchmarked option worth adding to a bake-off alongside whatever tools you're already comparing — the combination of highest pass rate, lowest cost per pass, and fastest solve time on the same benchmark is a genuinely strong combined signal, not just a headline number. As with any benchmark claim in this fast-moving space, the practical move is running your own representative tasks through it rather than adopting a tool purely off a leaderboard result — but this is a legitimate, verifiable reason to put MiniMax's Code CLI on that shortlist.
This is also a reminder that "open source" and "best benchmark result" increasingly overlap rather than trade off against each other in the coding-agent space — a pattern worth watching as more labs, Chinese and Western alike, release competitive tooling under open licenses rather than reserving their strongest results for closed, paid products.
Related on explainx.ai
- How to read AI benchmarks and not get fooled
- What happened to GPT-6 Astra? Why the hype died down
- What is harness engineering? Complete guide
- Top 10 harness engineering concepts
- Primary sources: FrontierHarness Eval GitHub PR #13 · IT之家
This post is sourced to the FrontierHarness Eval project's GitHub PR #13 and IT之家 coverage, both reflecting data as of approximately September 18, 2026. Benchmark leaderboards update frequently — verify the current standing directly in the source repository before citing specific figures elsewhere.