Skip to content
Bud's Wiki

Kimi K2: Open Agentic Intelligence

TL;DR

Section titled “TL;DR”

Moonshot’s original Kimi K2 release (July 2025) — a 1.04T-parameter Mixture-of-Experts open-weight LLM with 32B activated parameters, pre-trained on 15.5T tokens with the novel MuonClip optimizer (Muon + QK-Clip stability mechanism) reporting zero loss spikes across the full pre-training run. K2 is explicitly positioned as an agentic foundation model: large-scale synthetic tool-use data and a joint RL stage over real and synthetic environments are the post-training axis, not chat alignment. Ships both Kimi-K2-Base and Kimi-K2-Instruct on Hugging Face under an open license. Headline results: 65.8 SWE-Bench Verified pass@1 (71.6 with parallel TTC), 47.3 SWE-Bench Multilingual, 66.1 Tau2-Bench, 76.5 ACEBench (En), 53.7 LiveCodeBench v6, 49.5 AIME 2025, 75.1 GPQA-Diamond — placing K2 as SOTA-among-open-non-thinking models at release, particularly on coding and agentic axes.

Key claims

Section titled “Key claims”
  • MuonClip optimizer combines Muon’s token efficiency with a novel QK-Clip mechanism that addresses Muon’s known training instability at large scale, enabling stable 15.5T-token pre-training of a 1T MoE with zero loss spikes [§Pre-training].
  • Scale: 1.04 trillion total parameters / 32 billion activated parameters, MoE architecture, pre-trained on 15.5T tokens — the largest reported Muon-family training run at release [§Pre-training].
  • Token efficiency is positioned as the dominant scaling coefficient: given the limited supply of high-quality human text, K2’s pre-training thesis is that maximizing intelligence extracted per token (via Muon-class optimization + synthetic data) is the binding constraint, not raw token count [§Pre-training].
  • Post-training is multi-stage and agentic-first: large-scale synthetic tool-use trajectory generation followed by a joint RL stage in which the model improves through interactions with both real and synthetic environments [§Post-training].
  • SWE-Bench Verified 65.8 pass@1 with bash/editor tools, single-attempt patches, no test-time compute, 8K output budget (16K for Agentless) — and 71.6 with parallel TTC sampling and an internal scoring model [README, evaluation note].
  • SWE-Bench Multilingual 47.3 pass@1 under the same single-attempt, no-TTC conditions [README].
  • Tau2-Bench 66.1, ACEBench (En) 76.5 — sets SOTA among open non-thinking models on the tool-use / function-calling axis at release [§Results].
  • Math and reasoning despite the non-thinking setting: AIME 2025 49.5, GPQA-Diamond 75.1, OJBench 27.1, with avg@k evaluation on AIME/HMMT/CNMO/PolyMath-en/GPQA-Diamond/EvalPlus/Tau2 to stabilize reporting [§Results].
  • Two checkpoints released: Kimi-K2-Base (foundation, for fine-tuning) and Kimi-K2-Instruct (post-trained, drop-in chat + agentic). Both open-sourced on Hugging Face [§Release].
  • Multimodal and thinking-mode are explicitly out of scope for this release — K2 is a non-thinking, text-only agentic model; those axes are reserved for subsequent releases [Launch tweet].
  • MoE architecture has 384 experts at 1T total / 32B active (per third-party model cards consistent with the blog and HF model card) [§Pre-training].
  • Released artifacts package: blog post + HF weights (moonshotai/Kimi-K2-Base and moonshotai/Kimi-K2-Instruct) + GitHub code at MoonshotAI/Kimi-K2 + API at platform.moonshot.ai + product surface at Kimi.ai [§Release].

Method

Section titled “Method”

K2 is a transformer MoE with 1.04T total parameters, 32B activated per token, and 384 experts. The training-recipe contribution of this release is twofold and lives in the pre-training stack.

MuonClip optimizer. Muon (Newton-Schulz-orthogonalized momentum updates on 2D weights) is the token-efficient base. Muon has a known instability mode at large scale — the QK projections in attention layers can develop unbounded logit magnitudes that cause loss spikes. K2 addresses this with QK-Clip, a stability-enhancing mechanism applied to the Muon updates of QK projections specifically. The result is a 15.5T-token pre-training run on the 1T MoE with zero loss spikes — the headline robustness datapoint of the release.

Agentic post-training stack. Two stages: (1) a large-scale agentic data synthesis pipeline that generates structured tool-use trajectories (multi-step reasoning + long-horizon planning + tool invocation), and (2) a joint RL stage where K2 improves via interaction with real and synthetic environments, incorporating preferences and self-critique. Concrete pipeline details (synthesizer architecture, RL objective, environment composition) are sketched at the blog level; the arxiv tech report 2507.20534 covers them in more depth.

What is not in this release. Multimodal input, thinking-mode (reasoning chains over hidden CoT), and the Agent Swarm orchestration that later K2.5 / K2.6 introduce are explicitly out of scope here — K2 is the text-only non-thinking foundation that those subsequent releases continued-pretrain or scale on.

Results

Section titled “Results”

Headline non-thinking-cohort comparison (open-source numbers cited from the blog; closed-source comparisons from the linked technical report):

  • Agentic / tool use: Tau2-Bench 66.1, ACEBench (En) 76.5 — SOTA among open non-thinking models at release.
  • Coding: SWE-Bench Verified 65.8 (single-attempt) → 71.6 (parallel-TTC + internal scoring); SWE-Bench Multilingual 47.3; LiveCodeBench v6 53.7.
  • Math: AIME 2025 49.5 (avg@k); OJBench 27.1.
  • Knowledge: GPQA-Diamond 75.1.
  • Tau2-Bench 66.1, ACEBench-En 76.5, SWE-Bench Verified 65.8, SWE-Bench Multilingual 47.3 is the central four-number summary the blog uses to position K2 as a frontier open-source non-thinking agentic model.

All metrics except SWE-Bench Verified (Agentless) are evaluated with an 8K output token budget; Agentless uses 16K. Stability evaluation uses avg@k on AIME / HMMT / CNMO / PolyMath-en / GPQA-Diamond / EvalPlus / Tau2.

Why it’s interesting

Section titled “Why it’s interesting”

K2 is the foundational ancestor of the Kimi-K2.x lineage already on the wiki — Kimi K2.6: Advancing Open-Source Coding and its predecessor K2.5 both continue from Kimi-K2-Base, so this fills a hole in the wiki’s open-MoE timeline. It is also the first filed entry that introduces MuonClip — Muon’s QK-Clip stability extension — at 1T-scale, which makes it the strongest known datapoint for Training stability at scale when scaling Muon-family optimizers; sibling The Newton-Muon Optimizer and Muon is Not That Special: Random or Inverted Spectra Work Just as Well probe Muon’s mathematical and empirical properties at smaller scale, but K2 is the production-scale stability proof. For MoE Routing Design it’s the largest filed open-MoE release at 1T total / 32B active / 384 experts — directly comparable to ERNIE 4.5 Technical Report (Baidu, 300B-A47B), DeepSeek-V4 collection release (Flash + Pro, up to 1.6T) (1.6T), and HunyuanImage 3.0 Technical Report (80B-13B image MoE). For Agentic Software Engineering it is the K2.5 / K2.6 baseline — 65.8 SWE-Bench Verified is the starting line K2.5 (76.8) and K2.6 (80.2) build off, and the “agentic-first post-training” framing this page tracks originates here.

See also

Section titled “See also”