Kimi K2: Open Agentic Intelligence

Open-weight MoE models large enough to be useful for agentic software work are expensive to serve, and their serving cost is dominated by FFN weight traffic rather than by attention - so the MoE total/activated split is the deployment decision, not an architecture detail.

A 1-TRILLION-TOTAL / 32-BILLION-ACTIVATED-parameter MoE. The training contribution named here is MUONCLIP, which adds a QK-clip technique to the Muon optimizer to address training instability while keeping Muon's token efficiency; K2 was pre-trained on 15.5T tokens with zero loss spike, then post-trained through agentic data synthesis and a joint RL stage against real and synthetic environments. For a serving reader the relevant content is the 1T/32B ratio itself: activated parameters set the per-token FLOPs while TOTAL parameters set the weight bytes that dominate decode - which is the quantity the repo's decode-gemm roofline cares about, and the reason a 1T-total model is only tractable on high-memory parts or with offload.

Paper's own claim is an ACCURACY claim, not a serving claim: state-of-the-art among open non-thinking models, with 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, 47.3 on SWE-Bench Multilingual, 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, 27.1 on OJBench, all WITHOUT extended thinking. No throughput, TTFT, or memory-footprint figure is published in the abstract, so speedup_reported records the benchmark claim and the absence of any serving metric. Those benchmark numbers are vendor self-reported and unaudited; they are recorded because they are the model's stated capability, not as a verified performance claim.

VENUE: arxiv-preprint WITH venue_track 'unknown'. arXiv comment reads only 'tech report of Kimi K2, with minor updates' - a comment that names NO venue, so there was nothing to copy even if copying were acceptable. No proceedings entry and no OpenReview decision record for this title. WHY IT IS A GAP FOR AN INFERENCE REPO: Kimi Linear is present (this repo's own attention record) and Kimi Linear's sibling model, but the 1T/32B MoE is the largest open-weights model in the record set and it is the one whose DEPLOYMENT arithmetic is most consequential - total parameters dominate decode bytes while activated parameters dominate prefill FLOPs, and the ratio between those two is exactly the knob the repo's capacity-planning docs assume. Open weights ARE released (base and post-trained checkpoints), which is why adoption is in-upstream-engine rather than research-only. THE CONSUMPTION PROFILE is the practical caveat: an agentic model optimized for non-thinking inference under sustained tool use is the workload that maximises PREFIX REUSE across turns, which is the Preble / Cache-Craft / LMCache regime rather than the single-turn throughput regime most published benchmarks measure. VENUE RE-CHECKED 2026-10-05 (wave 6, agent w6-papers) - STILL NO VENUE, and the wave that could not verify this is now three waves old, so the negative is re-run rather than inherited. Wave 4 reached the same answer from the OpenReview index; this pass adds two corpora it did not sweep. ACL Anthology search for this title returns no paper page of its own - only unrelated PDFs that CITE it, which is the strongest available evidence that no ACL-family proceedings entry exists. Crossref bibliographic query for the exact title returns no proceedings-article or journal-article of this paper either; the hits are other papers on adjacent topics (a different elastic KV paper, a different memory-disaggregated caching system, a different concurrent-LoRA serving system), none of them this work. So the answer is arxiv-preprint with venue_track 'unknown', and it is a POSITIVE finding rather than an unresearched gap: this is a widely cited paper with no peer-reviewed venue, which is a fact about the literature that a reader needs to know. The claim that matters for a deployment reader is the one already in the record above - its consumption profile - agentic, non-thinking inference under sustained tool use - is the workload that maximises PREFIX REUSE across turns, so it belongs in the Preble / Cache-Craft / LMCache reading list regardless of venue. arxiv_id and every other field unchanged; only `updated` moved, because `updated` is the date the record was last verified against its sources.