Kimi K3 Release: A 2.8 Trillion Parameter Model for Agentic Coding

Kimi K3 Release: A 2.8 Trillion Parameter Model for Agentic Coding

Kimi K3 delivers frontier-level intelligence for agentic work

Kimi K3 is Moonshot AI's most capable model to date, designed specifically for agentic coding and complex knowledge work. In internal and third-party evaluations, K3 demonstrates performance that rivals the current industry leaders, ranking second only to Claude Fable 5 and GPT-5.6 Sol in overall intelligence.

Key benchmark achievements include:

  • GDPval-AA v2: Kimi K3 scored 1687, placing it ahead of Claude Opus 4.8 Max (1600) and behind only Claude Fable 5 Max and GPT-5.6 Sol Max.
  • AA-Briefcase: K3 scored 1527, ranking second overall, surpassing GPT-5.6 Sol Max (1495) and trailing only Claude Fable 5 Max. This benchmark specifically evaluates agentic capability in long-horizon knowledge work.

Massive scale and architectural efficiency

Kimi K3 is one of the largest open models available, featuring 2.8 trillion total parameters. To manage this scale while maintaining inference efficiency, Moonshot AI utilized the Stable LatentMoE (Mixture of Experts) framework.

This architecture significantly increases sparsity, activating only 16 out of 896 experts per token. This structural approach provides roughly 2.5x the scaling efficiency of its predecessor, Kimi K2, allowing the model to convert compute into capability more effectively. Based on the total parameter count and expert distribution, the model is estimated to have approximately 50 billion active parameters during inference.

Breakthroughs in autonomous chip design and compilers

Kimi K3 has demonstrated advanced long-horizon agentic capabilities through two highly technical proof-of-concept projects:

  1. Autonomous Chip Design: In a 48-hour autonomous run, K3 designed a chip to serve a nano model based on its own architecture. Using open-source EDA tools and the Nangate 45nm library, it produced a design within 4 mm² that closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation.
  2. GPU Compiler Development: The model reportedly coded an entire GPU compiler from scratch, with some GPU kernels outperforming Triton in specific instances.

Technical specifications and API constraints

Kimi K3 supports a 1 million token context window and includes native vision capabilities. However, current API implementations have several rigid constraints:

  • Reasoning Effort: Currently only supports "max" reasoning effort; thinking mode is always enabled by default.
  • Fixed Parameters: Temperature (1.0), top_p (0.95), and penalties (presence and frequency = 0) are fixed and cannot be modified in requests.
  • Vision Input: Public image URLs are not supported; users must use base64 encoding or specific file IDs.
  • Web Search: Currently being updated and is not recommended for production workflows.

Pricing and Market Positioning

Kimi K3 is positioned as a direct competitor to the Anthropic Sonnet series. Pricing is set at $3 per 1 million input tokens and $15 per 1 million output tokens, with a cached input rate of $0.3 per 1 million tokens.

While the raw pricing is competitive with frontier models, community feedback suggests that the actual cost-effectiveness depends on reasoning efficiency. Some users have noted that K3 can be "slow" and may spend a high volume of reasoning tokens on a single prompt, which can increase the effective cost per task compared to more efficient models like GPT-5.6 Sol.

Community Insights and Critiques

Users on Hacker News have shared mixed early experiences regarding the model's real-world utility:

"On the first try, Kimi K3 just found the source of a bug that Fable 5 hasn't been able to pinpoint in multiple attempts."

"I have just tried a single prompt in the web interface and it is still not finish reasoning. It thinks too much and often repeats the same stuff over and over."

Other concerns were raised regarding data privacy, as Moonshot AI's terms of service indicate that customer content may be used to improve their models unless a separate enterprise agreement is in place.

Sources