GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Moonshot AI Ships Kimi K3: 2.8T Open-Weight Multimodal Model at $3/$15 Per Million

Moonshot AI launched Kimi K3 on July 16, making it immediately available via API through providers including OpenRouter. The model is 2.8 trillion parameters, open-weight, and multimodal, built around an architecture that uses KDA (Key-value Decomposed Attention) and Attention Residuals for computational efficiency.

What K3 Is Built For

K3 targets the specific failure modes of frontier coding agents: navigating large repositories, chaining tool calls across multiple steps, debugging against images and runtime logs, and iterating across test and feedback cycles. The architecture choices — KDA and Attention Residuals — are explicitly aimed at reducing the KV-cache overhead that causes performance degradation on long agentic runs.

The 1M-token context window gives K3 room to hold entire codebases in context. That matters for repository-level tasks where retrieving the right file at step 40 of a 60-step agent run determines whether the job completes.

Pricing and Positioning

MetricKimi K3Kimi K2.7-CodeClaude Opus 4.8
Input ($/M)$3.00$4.00$15.00
Output ($/M)$15.00$4.00$75.00
Context1M1M200K
Parameters2.8Tundisclosedundisclosed

K3 positions as a frontier-class reasoning model at roughly one-fifth the output cost of Opus 4.8. That gap will be the benchmark story: if K3 matches Opus 4.8 on agentic coding benchmarks, the cost differential becomes a structural argument for open-weight models in enterprise agent deployments.

Benchmark Expectations

K3 was expected, before launch, to match or surpass Anthropic’s Opus 4.8 on frontier benchmarks — according to pre-launch reporting citing anonymous sources familiar with Moonshot’s internal evaluations. Specific SWE-bench Verified, Agent Arena, or Terminal-Bench numbers have not been published at launch.

The K2 line preceded K3 with strong credentials: Kimi K2.6 hit 80.2% on SWE-bench Verified and 96.0% on tau2-bench Telecom, putting it in Opus 4.7 territory on agentic evals. Kimi K2.7-Code scored 81% on MCP Mark Verified and was the first open-weight model accepted into GitHub Copilot’s model picker. K3 at 2.8T parameters is a substantially larger system.

The Open-Weight Equation

Kimi K3 joins a small group of frontier-scale open-weight models — alongside DeepSeek V4 Pro (1.6T params, Apache 2.0) and GLM-5.2 (744B at 2-bit quantisation) — that are compressing the performance gap with closed API models. The prior Kimi K2.7-Code was already in GitHub Copilot, meaning K3’s open weights land in a market with proven enterprise distribution.

Moonshot AI is expected to publish full evaluation numbers, including SWE-bench and agentic benchmark results, in the days following launch.