Moonshot's Kimi K3: Largest Open-Weight Model Ever

Moonshot AI released Kimi K3 — at 2.8 trillion parameters, the largest open-weight model ever published — announcing the architecture on July 16 and pushing the weights public on July 27. Independent evaluations put it trading with the top closed U.S. systems on coding and agentic workloads, a first for an open release at this scale.

Key facts

• 2.8T total parameters — the largest open-weight checkpoint ever; published under a modified MIT license on Hugging Face.
• Hybrid linear attention (Kimi Delta Attention + attention residuals): sub-quadratic attention that keeps long-context serving feasible at this scale.
• 1 million-token context window with native vision — images, video, and documents inside the context.
• Benchmarks on release: Artificial Analysis Elo of 1,547 — second overall on that index, ahead of several premium closed offerings in earlier rounds.
• API pricing: $3.00/M input, $15.00/M output — a step below the GPT-5.6 Sol tier (see our pricing briefing and the DeepSeek V4-Flash report for the emerging cost picture).
• Built for agents: Moonshot explicitly trained for long-horizon agentic work — tools, code, multi-step reasoning — rather than chat polish.

The announcement

Moonshot’s July 16 announcement positioned Kimi K3 as the answer to a question the Kimi line has chased for a year: can a frontier-scale model be released open, and can it be served at cost that starts a price war rather than a compute budget? The architecture covers the serving side — the hybrid-attention stack keeps inference tractable at 2.8T — and the weights release, the largest ever, covers the open side.

The July 27 weights drop made the claim concrete: any team with GPU capacity can now run a model whose benchmark profile sits next to major closed systems, at per-token prices those systems cannot touch.

Why it matters

• The open frontier is now real. Kimi K3, Alibaba’s Qwen3.8-Max (2.4T, open weights, now live), and DeepSeek V4-Flash have closed the gap in weeks, not years. Closed vendors are left defending price on alignment, latency, and trust — not raw capability alone.
• Serving economics flip. For long-context workloads, the hybrid-attention architecture changes the KV-cache and serving math: 1M-token context becomes a budget question rather than a GPU-acquisition project.
• Agent economics improve immediately. Frontier-quality reasoning at open-model prices changes the unit economics of Agent 2.0 — the same force behind the OpenAI Luna price cut.
• Ecosystem gravity. A 2.8T open base invites fine-tunes, adapters, and tooling in a way a closed API cannot — the same network effect that made the Qwen family the developer default this year.

What to watch

• Real serving costs. Independent clusters publishing $/M token at 2.8T with hybrid attention vs. closed API prices — the actual production number deserves triage.
• Fine-tune storm. Which agent frameworks adopt Kimi K3 as a flag-1 open backend, and how quickly.
• The closed response. Whether the closed tier responds with a “context-lite” cheaper tier to blunt the 1M-context person-economics pitch.

Official source

• Hugging Face: Kimi-K3 model card and weights
• Moonshot AI announcement (July 16): moonshot.ai/blog
• Artificial Analysis: official Kimi-K3 scorecard
• Shared-basis comparison: Qwen3.8-Max vs Kimi K3: The Numbers Beat

Updated August 8, 2026 — weights live since July 27.