Kimi K2.7-Code

PRELIMINARY / VENDOR-DATA-ONLY calibration. Moonshot AI, released 2026-06-12. Open-weight (Modified MIT, weights on HuggingFace), built ON TOP OF K2.6 (anchor) as a coding-focused successor. Pairs with Kimi Code (terminal-first coding agent, plans from $19/mo). CRITICAL CAVEAT: ALL performance numbers are Moonshot's OWN PROPRIETARY benchmarks — NO independent replication exists (VentureBeat reports practitioners dispute the gains; the three headline metrics are vendor-internal). This record is therefore PRELIMINARY and scored CONSERVATIVELY: it inherits K2.6's calibration with only coding/efficiency nudges, NOT a full re-score. ZERO eval snapshots wired — vendor-proprietary benchmarks (Kimi Code Bench v2, Program Bench, MLS Bench Lite) fail the traceable-independent-source requirement and are deliberately NOT created as Eval records. Vendor claims vs K2.6: +21.8% Kimi Code Bench v2, +11.0% Program Bench, +31.5% MLS Bench Lite, ~30% LOWER reasoning-token usage. DIMENSION RATIONALE: held at K2.6 levels across the board; the ONLY deltas from K2.6 are Speed 14 (UP from K2.6's 11 — the ~30% reasoning-token reduction directly addresses K2.6's verbosity/slowness, the one credible + architecturally-plausible improvement) and Creativity/Multimodal nudged down 1 (coding-specialization focus). Code Gen 17 + Planning 18 held = no independent basis to raise above K2.6. REVISIT when Artificial Analysis / SWE-Bench Pro / independent evals publish; expect to re-anchor then. Pricing not yet published (K2.6 ref: ~$0.67-0.95 in / $3.39-4.00 out per M). Sibling/successor to Kimi K2.6.