Provenance Record

Well-structured News piece (566 words, within the 400-1200 range). Clean Overview / What We Know / What We Don't Know / Analysis arc. Technical depth is appropriate and the efficiency framing matches the actual significance of the release.

All 4 source snapshots read from disk (status 200, none bot-blocked). source-0.html.gz (Hugging Face model card, PRIMARY): confirms verbatim 'reducing thinking-token usage by approximately 30% compared with Kimi K2.6', Total Parameters 1T, Activated 32B, 61 Layers, 384 Experts, 8 Selected per Token, 1 Shared Expert, 256K Context, Modified MIT License, 'Interleaved Thinking and Multi-Step Tool Call', and 'enabled by default and can't be disabled' (supports the forced-reasoning claim). Benchmark table confirms K2.7 Code 62.0 / GPT-5.5 69.0 / Claude Opus 4.8 67.4 on Kimi Code Bench v2 and 76.0 / 79.4 / 81.3 on MCP Atlas. source-1.html.gz (SiliconANGLE, K2.6, April): confirms 'Kimi-K2.6 ... 384 so-called experts ... it only uses eight experts to generate a response' and 1T parameters; supports the K2.6 same-footprint and April-release claims. source-2.html.gz (SiliconANGLE, K2.5, January): confirms 'open-source Kimi K2.5 model with 1T parameters', MoE with 1 trillion parameters, weights made available on Hugging Face. source-3.html.gz (VentureBeat): NOT bot-blocked in this capture (full 128KB article, status 200); confirms verbatim headline 'Kimi K2.7-Code cuts thinking tokens 30% — but practitioners say the benchmarks don't check out', the 21.8% / 31.5% gain figures, 'All three are proprietary benchmarks run by Moonshot AI', and practitioner skepticism about independent benchmarks.

Every specific traces to a cited source. CRITICAL Rule 9 satisfied: all architecture (1T/32B/384/8/61/256K) and the two head-to-head benchmark numbers (62.0/76.0 and competitor figures) are cited to the Hugging Face model card, not the secondary outlets. The 30% reduction figure and the 1T parameter count both appear verbatim in the model card. The 21.8% / 31.5% vendor-claimed gains are correctly attributed to VentureBeat (which itself attributes them to Moonshot). No headline or lead claim rests solely on VentureBeat — the headline/lead 30% claim and 1T descriptor both trace to the primary model card; VentureBeat carries only corroborated color and the practitioner-skepticism caveat. The article makes no specific release-date claim, avoiding an unsourced date. Internal cross-reference /article/2026-05/10-moonshot-ai-closes-2-billion-round... resolves to a real published article.

High-quality, fully sourced News submission. Signature and hash valid, all 18 automated checks pass, all four source snapshots personally read and confirm the claims attributed to them. Architecture and benchmark specifics correctly trace to the primary Hugging Face model card; vendor claims are framed as vendor claims. APPROVE.