Moonshot AI

Type: AI Research Lab (Beijing) — open-weight frontier models

Moonshot AI enters this knowledge base with Kimi K3 (released 2026-07-16), a 2.8-trillion-parameter mixture-of-experts model described in coverage as the largest open-weight model to date. Its significance here is not size but competitive posture: where DeepSeek V4 (April 2026) attacked the closed frontier on cost — near-frontier capability at roughly one-sixth the price — K3 attacks it on capability while raising price, launching at Claude Sonnet-tier rates and becoming the most expensive model a Chinese lab has shipped to date. That is a genuinely different strategy at the open-weight frontier, and it is the reason this entity is worth tracking separately from DeepSeek.

Architecture: MoE with ~2.8T total parameters, 896 experts, 16 active per token (~1.8% activation) — forward-pass compute is far below what the headline count implies. Moonshot names the architecture "LatentMoE" and cites two new attention mechanisms, Kimi Delta Attention and Attention Residuals, aimed at compute efficiency and long, low-oversight coding/agent sessions. Context window is 1,000,000 tokens with native vision input.

Evidence caveat — read this before quoting any number below. Moonshot's own blog did not resolve to announcement content at ingest time. Everything here comes from an independent technical review (Simon Willison), wire coverage quoting Moonshot's press release, and third-party benchmark aggregators. Open weights were not yet released at ingest (promised 2026-07-27), so no benchmark claim on this page has been independently reproduced. This is an analysis-grade source describing vendor claims.

Pricing and the posture reversal

A ~3× input and ~3.75× output price increase against its own prior generation, landing on the same tier as Anthropic's Claude Sonnet family. See Inference & Token Economics for why this matters: it is a live counter-example to the assumption that open-weight entrants compete by undercutting.

Benchmark claims (all vendor or third-party reported, unreproduced)

• Moonshot's own evaluation: K3 trails Claude Fable 5 and OpenAI's GPT-5.6 Sol on overall performance, but "mostly beats" Claude Opus 4.8 and GPT-5.5, and substantially outperforms other tested models on coding/agentic tasks.
• Arena.ai Frontend Code leaderboard: ranked #1, ahead of Claude Fable 5 — reported as the first time an open-weight model has topped it.
• Vercel professional web-engineering evaluation: ranked #1 overall.
• Artificial Analysis (long-horizon knowledge work Elo): 1547, +732 over K2.6, "behind only Claude Fable 5"; ~21% fewer output tokens than K2.6 on the same eval; ~$0.94 cost per task.

Internal conflict worth preserving: the vendor's own eval places K3 behind Fable 5 overall, while a task-specific leaderboard places it ahead of Fable 5. Both figures come from the same source. These are not reconcilable into a single ranking — they measure different things, and the vendor has an incentive to cite whichever favours it. Do not collapse them.

Key Contributions

• Kimi K3 (2026-07-16): ~2.8T-parameter MoE, 896 experts / 16 active per token, 1M context, native vision — described as the largest open-weight model to date; open weights due 2026-07-27 (Kimi K3)
• "LatentMoE" architecture with Kimi Delta Attention and Attention Residuals, targeted at long, low-oversight agentic coding sessions (Kimi K3)
• A capability-led (not cost-led) open-weight posture: first Chinese-lab release priced at closed-frontier mid-tier rates, up sharply from its own prior generation (Kimi K3)

Observed operational quirks

• Single reasoning-effort setting ("max") at launch, with no lower tier — token consumption on trivial prompts is high (an observed 13,241 reasoning tokens, ~$0.25, on a hello-world-grade request). See Chain-of-Thought Reasoning on deliberation budgets moving out of buyer control.
• ~85-token implicit system prompt inferred from a 10-word prompt billing as 95 tokens — real per-request overhead that headline pricing does not capture.
• The reviewer's own caveat: his pelican-SVG test is a smoke test, explicitly not a capability measure, and should not be used comparatively.

Mentioned In

• Inference & Token Economics — the July-2026 three-lab pricing divergence
• Chain-of-Thought Reasoning — non-configurable reasoning effort

Related Entities

• DeepSeek — the other open-weight frontier lab in this KB; opposite pricing posture
• Anthropic — price-tier peer (Sonnet) and the benchmark comparator K3 is measured against
• OpenAI — closed frontier peer (GPT-5.6 Sol)

Outstanding Questions

• Do the benchmark claims survive the 2026-07-27 open-weight release and independent reproduction?
• Is the price increase a confidence signal, a serving-cost pass-through (2.8T params, even at 1.8% activation, is expensive to host), or both?
• Does a lower reasoning-effort tier ship, and what does that do to the ~$0.94-per-task figure?

Changelog

• 2026-07-22 — Created from the Kimi K3 analysis source. Records architecture, the K2.6→K3 price increase, all vendor/third-party benchmark claims with unreproduced status, and the vendor-eval vs. Arena.ai internal conflict.