Kimi K3 is a 2.8T Open-Weight MoE Priced Like Sonnet 5 — and the Weights Ship in a Week

by Persephone

Moonshot's Kimi K3 — 2.8T-param MoE with 16 of 896 experts active per token and a 1M-token context window — launches closed at $3 / $15 per Mtok, identical to Claude Sonnet 5's introductory price. Open weights ship by July 27 under Modified MIT, the first 3T-class model to release with open weights.

Kimi K3 is a 2.8T Open-Weight MoE Priced Like Sonnet 5 — and the Weights Ship in a Week

July 20, 2026

Moonshot AI launched Kimi K3 at the World Artificial Intelligence Conference in Shanghai on Friday: 2.8T total parameters, 16 of 896 experts active per token, a 1M-token context window, priced at $3 per million input tokens and $15 per million output tokens. Weights ship by July 27 under a Modified MIT license, making K3 the first 3T-class model available with open weights (Kimi blog, 2026-07-16). The price parity with Claude Sonnet 5’s introductory API rate is the news that matters, not the parameter count.

The Architecture: 2.8T MoE with Two New Tricks

K3 is a sparse mixture-of-experts transformer with 2.8T total parameters and 896 expert FFNs per layer. Sixteen experts activate per token — about 1.8% of the experts, yielding roughly 50B activated parameters per forward pass (Kimi blog, 2026-07-16). That is the operational footprint: K3 routes like a 50B dense model but carries the memory cost of a 2.8T-parameter MoE.

Two architectural elements are new. Kimi Delta Attention (KDA) replaces standard multi-head attention with a delta-rule update that compresses recent-token attention state into a smaller cache footprint. Attention Residuals (AttnRes) carry attention-derived signals across MoE layers as auxiliary residuals alongside the FFN output, instead of relying solely on the FFN pathway. Moonshot’s published scaling-efficiency analysis claims roughly a 2.5x improvement in compute-to-quality ratio versus Kimi K2 on long-horizon reasoning, attributed to KDA reducing per-token attention cost and AttnRes improving gradient signal at depth (Kimi blog, 2026-07-16). Both mechanisms are documented in the technical report Moonshot published alongside the launch.

Context length lands at 1M tokens. Native multimodal input handles text and image together; audio and video are scheduled for a later checkpoint.

Pricing: Parity with Sonnet 5

Kimi K3 ships at $3 per million input tokens and $15 per million output tokens — identical to Claude Sonnet 5’s introductory API price (Anthropic pricing, 2026-07-20). It is the highest launch price any Chinese frontier lab has set for a flagship model (Simon Willison, 2026-07-16).

The strategic read is straightforward. DeepSeek, Qwen, and Zhipu have anchored their flagship API pricing 60–80% below US frontier equivalents. Moonshot is not playing that game with K3. They are pricing at parity and betting that quality plus the open-weights release next week carries adoption. The bet is a real product-category shift: $3 / $15 with a Modified MIT license in a week is a different category from $0.40 / $1.20 under Apache 2.0 today.

Quality: Artificial Analysis Long-Horizon K3 = Elo 1547

Artificial Analysis’s long-horizon evaluation ranks Kimi K3 at Elo 1547, second only to Claude Fable 5 (Elo 1571) and ahead of GPT-5.6 Pro (1543) and Opus 4.8 (1538). Cost per task on the long-horizon benchmark lands at $0.94 — cheaper than GPT-5.6 Sol ($1.04 per task) and less than half of Opus 4.8 ($1.80 per task) (Simon Willison citing Artificial Analysis, 2026-07-16).

The K2.6 → K3 jump is +732 Elo on long-horizon knowledge work. That is a generational move inside a single model family, and it places K3 in the same operational tier as the top two closed US models on the workload Artificial Analysis measures.

Market Reaction: Zhipu -27%, MiniMax -16%

The pricing posture landed in Hong Kong the same morning. Zhipu AI closed Friday down 27% on turnover more than four times the 30-day average. MiniMax closed down 16% on roughly double normal volume (BBC News, 2026-07-18). Investors read the launch as Moonshot resetting the price-quality frontier, with the open-weights release next week as the second shoe.

Moonshot itself is unlisted. It is backed by Alibaba and Tencent and has reportedly been weighing a Hong Kong IPO at a target valuation between $20B and $30B (BBC News, 2026-07-18). The Friday reaction is the first time an open-weights launch from a Chinese lab has moved both US-listed and Hong Kong–listed competitors on a single event.

How You Can Use It Today — and What Changes July 27

Kimi K3 is live right now in closed beta across four surfaces: kimi.com chat, Kimi Work (the productivity surface), Kimi Code (the IDE assistant), and the Kimi API. Signup is gated behind a waitlist; existing Moonshot accounts have been auto-approved in waves (Kimi blog, 2026-07-16).

Starting July 27, the weights ship under Modified MIT — a license similar to standard MIT but with two carve-outs: a 100M-MAU threshold above which commercial deployment requires a separate license from Moonshot, and an explicit acceptable-use clause tied to Chinese AI regulation. The license structure is documented at the community mirror (kimik3.xyz). For most platform teams the practical read is: weights are usable for self-hosted research, internal tooling, and commercial products below the MAU threshold; anything targeting consumer-scale end users is a separate conversation with Moonshot.

Self-hosting 2.8T parameters is not a laptop story. With 16 experts active per token, the practical FP8 inference footprint lands in the 400GB–500GB range for weights plus KV cache — within reach of a single 8x H200 node or a multi-GPU consumer rig with NVLink, but well outside the LM Studio bracket. The realistic paths for most teams are: (1) Kimi API at $3 / $15 for production, (2) vLLM or SGLang self-host on rented H200 capacity, or (3) a quantized checkpoint (INT4 or AWQ) when community quantizations land in the days after the open release.

What to Watch

Three things over the next 10 days.

1. The Modified MIT license text. The 100M MAU threshold and acceptable-use carve-out need to be read carefully before any consumer product decision. If Moonshot has written it the way early coverage suggests, the license is more open than Llama and less open than Apache. The exact wording matters (kimik3.xyz).

2. Community quantization. GPTQ, AWQ, and GGUF checkpoints will land within 48–72 hours of the weight release if Moonshot’s track record with K2 holds. The INT4 checkpoints are what make the model usable on single-H100 or dual-A100 rigs without the 400GB+ FP8 footprint.

3. The Hong Kong IPO. If Moonshot prices inside the $20B–$30B range after the open-weights launch, it sets the valuation reference for every Chinese open-weights lab going forward. Zhipu, DeepSeek, and MiniMax all have IPO files pending; the read-through will be immediate.

The bigger story is structural. Kimi K3 is the first open-weights model that lands in the same quality tier as the closed US frontier on a benchmark Artificial Analysis publishes with methodology anyone can audit — and it is priced at parity with the model it competes with on that benchmark. The “open weights mean cheap and slightly behind” framing that held from Llama 2 through DeepSeek V3 just stopped being accurate.

Sources

Prior TopClanker Coverage