Moonshot AI's Kimi K3 Beats Claude at Coding for 4.6x Less Money
Kimi K3 debuts at #3 on DeepSWE, matching Claude Fable and GPT-5.6 Sol as the first open-weights model to reach frontier-level coding performance at 2.8 trillion parameters.
- Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weights model, debuted at #3 on the DeepSWE leaderboard with a 69% score, just behind Claude Fable 5 (70%) and GPT-5.6 Sol (73%).
- It is the largest open-weight model ever announced, surpassing DeepSeek (1.6T), and introduces three new architectural components: Kimi Delta Attention (KDA), Attention Residuals, and Stable LatentMoE.
- KDA enables up to 6.3x faster decoding in million-token contexts; AttnRes adds ~25% training efficiency at under 2% extra compute cost.
- API is live now at $3/$15 per million input/output tokens -- roughly 3-5x cheaper than comparable closed-model alternatives; full weights drop July 27 on Hugging Face.
- DeepSWE, the benchmark behind the headline, is a contamination-free, long-horizon coding benchmark spanning 91 repos across 5 languages, designed to separate models that cluster on SWE-Bench.
- Self-hosting requires 64+ accelerators; most teams will access K3 via the Kimi API or managed inference providers once weights ship.
Moonshot AI just dropped what may be the most consequential open-weights model release since DeepSeek R1. Kimi K3 is a 2.8-trillion-parameter sparse Mixture-of-Experts model that debuted at #3 on the DeepSWE leaderboard, scoring 69% on a benchmark designed specifically to stress-test frontier coding agents on long-horizon, real-world software engineering tasks. That puts it within a few percentage points of Claude Fable 5 (70%) and GPT-5.6 Sol (73%) -- the current top two -- and ahead of every other model on the list.
The benchmark that caught everyone's attention
DeepSWE, built by Datacurve, is a coding benchmark designed to avoid the saturation problem plaguing existing leaderboards. Today's leading public coding benchmarks are starting to saturate at the frontier, with top models clustering within a narrow score band. DeepSWE is built to separate them. The key design decisions that make it harder to game:
- Contamination-free: Tasks are written from scratch, not adapted from existing commits or PRs, so no model has seen the solution during pretraining.
- Real-world complexity: Prompts are roughly half the length of SWE-bench Pro's, yet solutions require 5.5x more code and ~2x more output tokens.
- Reliable verification: Verifiers are hand-written to test software behavior rather than implementation details.
- High diversity: Tasks span a broad pool of 91 repositories across 5 languages.
The benchmark also found serious problems with the status quo. SWE-Bench Pro, the current leading agentic coding benchmark, has a verifier that misgrades agent outputs at rates of 8% false positives and 24% false negatives. On DeepSWE, those rates drop to 0.3% and 1.1% respectively -- a meaningful improvement for anyone trying to make real decisions about which model to use.
2.8 trillion parameters, but only 16 are active at a time
Kimi K3 reports 2.8 trillion total parameters. It is a sparse Mixture-of-Experts model, routing each token through a small subset -- 16 of 896 routed experts -- rather than using all parameters densely for every token. Think of it like a massive panel of specialists where only a handful are consulted per question. This is what makes a model this large computationally tractable to serve.