Moonshot AI opens Kimi K3 with 2.8 trillion parameters, challenging rivals

Moonshot AI's decision to open-source a 2.8-trillion-parameter model that matches proprietary rivals on key benchmarks marks a turning point in the economics of artificial intelligence.

Moonshot AI released the full model weights for Kimi K3 on July 27, giving developers free access to a 2.8-trillion-parameter system that matches Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on browsing and coding benchmarks at a fraction of the cost. The Beijing-based company also open-sourced the infrastructure technologies — MoonEP, FlashKDA and AgentEnv — that underpin the model's training and inference.

"Kimi K3 is our most capable model," Moonshot AI said in a statement. The model uses a Mixture-of-Experts architecture with 896 experts, activating 16 per inference to maintain computational efficiency. On the BrowseComp benchmark, which evaluates real-world research and browsing, K3 scored 91.2, compared with Claude Fable 5 at 88.0 and GPT-5.6 Sol at 90.4. The cost differential is larger: a single coding rollout costs $4.65 for K3 versus $13.41 for Claude Fable 5, delivering 2.8 times more solved tasks per dollar. At a full 452-rollout sweep, the total cost is $2,103 compared with $6,010.

The open-sourcing of a frontier-class model threatens the pricing power of proprietary leaders. Chinese AI models now account for about 60% of token usage by US companies on OpenRouter, the API marketplace. Nvidia Chief Executive Officer Jensen Huang said US companies "absolutely should be allowed to use Chinese models," a statement that carries weight given Nvidia's role as the primary hardware supplier for AI workloads.

Agent Swarm Architecture Drives Parallel Execution

Kimi K3 inherits the "Agent Swarm" framework introduced in Kimi K2.5, which replaces sequential tool-call execution with parallel task delegation. The system can deploy up to 300 sub-agents per task, managing more than 4,000 tool calls and running 4.5 times faster than single-agent sequential execution, according to Moonshot's technical paper. This architecture transforms task completion time from a function of sequential steps to a function of parallel processing capacity — a structural advantage for complex software engineering projects that require simultaneous research, design and development.

The model also incorporates native vision understanding through a three-dimensional vision encoder integrated during pretraining, a capability most open-weight competitors lack. DeepSeek's current open-source offerings remain text-centered, while K3's vision capability enables visual-to-code generation and real-world software engineering tasks that require interpreting screenshots and layouts. The model implements Kimi Delta Attention, a hybrid linear attention mechanism that provides a 6.3 times decoding speedup in million-token contexts, supporting a 1-million-token context window.

Enterprise Adoption and the Open-Weight Calculus

US companies are already integrating Kimi models into production. DoorDash has adopted Kimi for lower-level internal tasks while reserving Anthropic's models for higher-level functions. Coinbase confirmed internal use. Cursor, the coding startup acquired by SpaceX, built its product on a Kimi foundation. The adoption data suggests that cost-sensitive enterprises are treating open-weight models as a viable alternative to per-token pricing from US labs.

The competitive pressure extends beyond Moonshot. Thinking Machines released an open-weight model called Inkling last week, and Alibaba's Qwen team has signaled the imminent launch of Qwen 3.8. The proliferation of capable open-weight models is squeezing the pricing power of OpenAI and Anthropic, whose recent releases have shown only incremental benchmark improvements.

The structural economics favor open distribution. Per-inference pricing creates a variable cost that scales with usage, disincentivizing high-volume or low-margin applications. Open weights eliminate that friction: distribution carries zero marginal cost per copy, modifications require no external permissions, and deployment can occur on local hardware or air-gapped networks without per-token accounting. Box Chief Executive Officer Aaron Levie argued that falling token costs will drive up inference demand rather than shrink AI spending, benefiting infrastructure providers. For investors, the question is whether proprietary model vendors can maintain premium pricing as open-weight alternatives close the performance gap. Nvidia, as the primary beneficiary of any increase in inference compute demand, stands to gain regardless of which model architecture wins.

This article is for informational purposes only and does not constitute investment advice.