Kimi K3 is the biggest open model ever. It's still not the best.

Kimi K3 is the largest open model yet, at 2.8 trillion parameters, built by Moonshot AI with a novel attention architecture and a 1-million-token context window. But one number in its benchmark table stands out more: overall performance still lags behind Claude Fable 5 and GPT 5.6 Sol, with Claude Opus 5 recently matching Fable-level performance at half the cost, according to Claude Opus 5's benchmark results. That gap matters, not because K3 is weak (it isn't), but because it complicates the narrative that open models are catching up.

Where K3 wins

K3 leads on several individual benchmarks. On SWE Marathon, it scores 59.7, ahead of Claude Fable 5's 57.4 and GPT 5.6 Sol's 56.5. On Terminal-Bench 2.1, K3 hits 45.8, beating all proprietary rivals. BrowseComp's 1M-context variant gives it an 89.5, again above the competition, as a benchmark-by-benchmark breakdown confirms. These scores point to real strengths in long-horizon coding, autonomous tool use, and deep web research. In agentic tasks that demand sustained reasoning and interaction with environments, K3 often leads, reflecting advances in how models handle sustained interaction, an area where techniques like SEED distillation have shown promise.

But on comprehensive benchmarks like DeepSWE, Program Bench, and KCB 2.0, K3 sits behind the best proprietary models by several points, a pattern consistent with the observation that even top agents score 49% on complex classification tasks, as shown by the HSCodeComp benchmark. The overall picture is a scatter of wins and losses, not a clear victory.

What the architecture tells us

K3's technical story is more interesting than the headline numbers. Kimi Delta Attention and Attention Residuals are genuine innovations designed to handle deep networks and long sequences without the usual degradation. Stable LatentMoE activates 16 out of 896 experts, and the training methods (Quantile Balancing, Per-Head Muon, quantization-aware training with MXFP4) address real scaling pains that earlier large models hit hard. These are engineering breakthroughs that could trickle down to other open-source projects, even if K3 itself doesn't top every leaderboard.

Pricing as a strategy

Moonshot AI is pricing aggressively. The Kimi API charges $0.30 per million tokens for cache-hit input, $3.00 for cache-miss input, and $15.00 for output. That undercuts comparable tiers from Anthropic and OpenAI. The company claims a cache hit rate above 90% on coding workloads, thanks to Mooncake's disaggregated inference architecture. The pricing relies on efficient inference, a focus of recent work like Together AI's open inference math. Combined with the open weight release scheduled for July 27, 2026, the pricing makes K3 attractive for organizations that want a 2.8T-parameter model they can inspect and host themselves.

What it means for the open-source frontier

The field has been sold a narrative: open models are catching up. K3 challenges that story head-on. It is the largest open model by a wide margin, yet it does not consistently beat models with fewer parameters but better data and post-training. That suggests the gap is not about compute alone. Data curation, alignment techniques, and proprietary training pipelines still give closed models an edge that raw scale cannot erase. The narrative is further complicated by the fact that open model demand is soaring, with Ollama suspending its premium plan due to traffic, and a coalition of 41 companies arguing that open-weight AI is America's best bet. Meanwhile, open-source models from the East are eroding the US advantage, as seen in the trend from the East.

K3 is a serious achievement. It proves open models can hit 3T-class scale and still function reliably, a nontrivial engineering feat. But the benchmark results also warn against assuming parameter count is a proxy for intelligence. The frontier is more than a number.