Provenance Record

Well-structured News piece with clear Overview / What We Know / What We Don't Know sections. Tone is neutral and hedged appropriately (Moonshot 'claims', 'positions', 'says'). Coverage is proportionate to a model-launch story: architecture, benchmarks, long-horizon agent demos, licensing terms, and honest acknowledgement of unverified internal runs. Word count 685 is within the News range (400-1200).

Four sources cited. Verified against local source snapshots in sources/2026-04/moonshot-ai-open-sources-kimi-k26-a-trillion-parameter-model-that-runs-300-agent-swarms-for-hours/ (all fetched 200 OK) and, where possible, via live WebFetch. (1) https://www.scmp.com/tech/big-tech/article/3350887/moonshot-ai-releases-flagship-model-open-source-push-continues — SUPPORTED. Confirms Moonshot's April 21 unveiling, the direct quote 'performed on par with or better than leading closed-source systems such as GPT-5.4, Claude Opus 4.6 and Gemini 3.1 Pro across several benchmarks', the 'independent verification of these claims remains limited, highlighting the lack of standardised evaluation' caveat, and Alibaba Cloud's release of Qwen3.6-Max-Preview as a closed model on the same Monday. (2) https://siliconangle.com/2026/04/20/moonshot-ai-releases-kimi-k2-6-model-1t-parameters-attention-optimizations/ — SUPPORTED. Confirms the April 20 release date, MoE architecture with 384 experts and 8 activated per inference, 400M-parameter vision encoder, multi-head latent attention (MLA) as a hardware-efficient attention variant, SwiGLU activation, and Hugging Face availability. 'Kimi Code' CLI and Kimi chat app distribution is consistent with the headline/title and is also directly supported by the VentureBeat snapshot. (3) https://venturebeat.com/orchestration/kimi-k2-6-runs-agents-for-days-and-exposes-the-limits-of-enterprise-orchestration — PARTIALLY SUPPORTED. Directly confirms the quoted phrase '300 sub-agents executing across 4,000 coordinated steps simultaneously', the five-day autonomous run handling monitoring and incident response, the SysY compiler claim (10 hours, equivalent to four engineers over two months, 140 functional tests passed without human intervention), and the framing that long-horizon agents are 'exposing a critical gap in orchestration'. However, VentureBeat does NOT contain the comparison figures 57.7 (GPT-5.4), 53.4 (Claude Opus 4.6), or 54.2 (Gemini 3.1 Pro), nor the qualifier phrases 'high-compute setting', 'maximum effort', or 'thinking mode'. Those three numbers and phrasings were not found anywhere in any of the four source snapshots. The attribution is therefore unsupported. (4) https://the-decoder.com/open-weight-kimi-k2-6-takes-on-gpt-5-4-and-claude-opus-4-6-with-agent-swarms/ — SUPPORTED. Confirms the K2.6 scores 54.0 on HLE with Tools, 58.6 on SWE-Bench Pro, 83.2 on BrowseComp; the phrase 'falls behind on pure reasoning and vision'; the '300 sub-agents at once, each taking 4,000 steps' swarm topology; continuous execution of more than twelve hours in Rust, Go, and Python; and the modified MIT license with the 100M MAU / $20M monthly revenue attribution threshold.

All claims verified against at least one snapshot EXCEPT the competitor SWE-Bench Pro comparison numbers (57.7 / 53.4 / 54.2) and the effort-tier qualifiers (high-compute / maximum effort / thinking mode) attributed to VentureBeat. Those figures are not present in any of the four cited sources and must either be removed or replaced with a verifiable citation (e.g., the Moonshot blog post at kimi.com/blog/kimi-k2-6, which VentureBeat links to but this submission does not cite). Everything else — release dates, architecture details, licensing terms, agent-swarm specs, SCMP and The Decoder quotes, the Alibaba Qwen3.6-Max-Preview same-day context — maps correctly to the cited sources.

High-quality, well-sourced submission on a genuinely new topic, but one fabricated attribution (the SWE-Bench Pro competitor figures cited to VentureBeat) must be corrected before publication. This is a REQUEST_CHANGES, not a REJECT: the problem is localized to one sentence, the rest of the article is rigorously sourced, and a small rewrite with correct attribution (or removal) fixes it cleanly.