Source: LLM Stats — 2026-07-17
Summary
Moonshot AI's newly released Kimi K3 (in max-reasoning mode) scored 88.3% on Terminal-Bench 2.1, landing #2 overall and #1 among open-weight models — just 0.5 points behind GPT-5.6 Sol's leading 88.8%. Meta's Muse Spark 1.1 also joined the board this week at 80.0%, making this one of the fastest-moving stretches yet for the open-vs-closed gap on real terminal-agent tasks.
Key Takeaways
- Terminal-Bench 2.1 grades agents on completing real command-line tasks end-to-end, not multiple-choice or code-snippet correctness, so it's a harder proxy for "can this model actually operate a terminal" than most coding leaderboards.
- Kimi K3 closed the open/closed gap to half a point within days of its own release — a 2.8-trillion-parameter open MoE model trading blows with the current closed frontier.
- Meta's Muse Spark 1.1 entering at 80.0% shows the leaderboard churn isn't just a two-horse race between Moonshot and OpenAI.
- For teams choosing a model for agentic CLI work, the practical delta between the top open-weight and top closed option on this benchmark is now small enough that licensing, cost, and self-hosting constraints can reasonably decide the call.
Reel Script
Hook (16s / 36 words) An open-weight model just came within half a point of beating the best closed model in the world at operating a real terminal. That gap used to be measured in double digits.
Core Concept (55s / 120 words) Terminal-Bench 2.1 is a benchmark that doesn't ask a model to just write code — it drops an agent into an actual terminal and scores whether it can complete real, multi-step command-line tasks: install things, debug a broken build, chase down an error across files. Think of it as a driving test instead of a written exam — you actually have to operate the vehicle. This week Moonshot AI's Kimi K3, running in its highest reasoning mode, scored 88.3%. GPT-5.6 Sol, OpenAI's current best, scored 88.8%. That's a 0.5-point gap between a model anyone can download and run themselves, and the best closed model money can buy.
Hands-On (50s / 110 words) The numbers, stacked up: GPT-5.6 Sol at 88.8%, Kimi K3 max-reasoning at 88.3%, and Meta's freshly-added Muse Spark 1.1 at 80.0%. Kimi K3 is a 2.8-trillion-parameter open mixture-of-experts model — meaning only a fraction of those parameters activate per task, which is part of how it stays fast enough to compete on an agentic benchmark like this instead of just a static Q&A test. What's notable isn't just the score, it's the timing: K3 only shipped days before landing at #2, meaning Moonshot tuned it for exactly this kind of long-horizon, tool-using task from the start rather than bolting agent behavior on after the fact.
Takeaway (22s / 48 words) If your agentic tooling is locked to a closed frontier model purely for capability reasons, that justification is getting thinner by the week. Go check whether an open-weight option now clears your actual bar — not last quarter's bar.