On July 16, 2026, Moonshot AI released Kimi K3 — a 2.8 trillion-parameter mixture-of-experts model that Moonshot claims performs competitively with Fable 5 and “substantially outperforms” Opus 4.8 and GPT-5.6 Sol on agentic tasks. It’s the largest open-weight LLM released to date, and it debuted at #3 on the Artificial Analysis leaderboard, behind only Fable 5 and GPT-5.6 Sol.
What K3 actually is
K3 is a mixture-of-experts model with 896 routed experts, of which only 16 activate per token — the sparsity is what keeps a 2.8T-parameter model tractable to run at all. Moonshot says this routing setup delivers a 2.5x scalability improvement over K2, the previous generation.
The bigger architectural story is Kimi Delta Attention (KDA), a new attention mechanism built specifically to make the model’s 1,048,576-token (1M) context window usable rather than just advertised. Long-context windows are cheap to claim and expensive to make performant — KDA is Moonshot’s answer to the quadratic cost of attending over a million tokens. The model also has native multimodal (visual) understanding built in, rather than bolted on as a separate vision tower.
Where it actually wins: agentic coding
The headline benchmark numbers are the usual leaderboard game, but the agentic coding results are the more interesting signal for anyone building with these models day to day:
- 67.5 on DeepSWE — a benchmark for sustained software-engineering ability across a session, not single-shot code generation
- 77.8 pass rate on ProgramBench
- First place in the Frontend Code Arena at 1,679 points, ahead of Fable 5 in blind developer testing
DeepSWE and ProgramBench specifically test multi-step, tool-using workflows — a model that plans, executes, checks its own output, and adapts over an extended session, which is the actual shape of how coding agents get used in production now, not the isolated single-function benchmarks that used to define “coding models.”
Pricing: not the usual Chinese-lab discount
K3 is priced at $3 per million input tokens and $15 per million output tokens (cached input tokens run $0.30/M). That puts it at parity with Anthropic’s Claude Sonnet tier — and makes it the most expensive model a Chinese AI lab has shipped to date. Previous open releases out of China (DeepSeek, Qwen, K2 itself) leaned hard on being dramatically cheaper than their Western counterparts as the main pitch. K3 is the first to say, in effect, “match us on price because we’re worth it,” which is as much a signal of confidence as the benchmark scores are.
The open-weight caveat
Moonshot has committed to publishing full weights on Hugging Face by July 27, 2026, under a modified MIT license that permits commercial use. As of the model’s announcement, though, the Hugging Face org still only listed K2-series checkpoints — no K3 weights yet. Training data and full training code aren’t part of the release either, so “open weight” here means inference-ready checkpoints under a permissive license, not full training reproducibility. Worth rechecking after the 27th rather than assuming the release lands exactly on schedule.
The bigger picture
Every few months another open-weight release out of China gets framed as “closing the gap” with the US frontier labs, and the framing is getting less speculative each time. What’s notable about K3 specifically isn’t that it’s big — 2.8T parameters is more a statement than a practical deployment target for most teams — it’s that the gains are concentrated exactly where production agent workloads live: long-context handling via KDA, and sustained multi-step coding performance rather than single-turn benchmark chasing. Combined with a Western-tier price point, K3 reads less like a cost play and more like Moonshot betting it can compete on capability alone. Whether that holds up once the weights are actually downloadable and third parties can reproduce the agentic benchmarks independently is the thing worth watching next.
Sources: Fortune, Bloomberg, TechCrunch, CNBC, Tom’s Hardware.