Moonshot AI’s Kimi K3 represents a significant leap in open-weight AI model development. Released on July 16, 2026, this 2.8-trillion-parameter reasoning model has quickly become a focal point for developers seeking frontier-level intelligence with the flexibility of open weights. This analysis evaluates Kimi K3’s technical specifications, benchmark performance, and compares the leading API providers offering access to the model.

At a Glance

Metric	Value	Context
Intelligence Index Score	57	Well above median 32 for open-weight models
Context Window	1M tokens	~1,573 A4 pages equivalent
Input Modalities	Text + Image	Multimodal reasoning model
Architecture	MoE (2.8T total / 104B active)	16 of 896 experts activated per token
Input Pricing	$3.00/M tokens	Median: $1.75
Output Pricing	$15.00/M tokens	Median: $10.00
Cache Hit Price	$0.30/M tokens	90% discount
Output Speed	32.0 t/s	Below median for reasoning models of this class
TTFT Latency	144.34s	High relative to median due to always-on max reasoning
Output Tokens (Intelligence Index)	130M	Median: 63M
Evaluation Cost	$2,437.41	Full Intelligence Index run

What is the Kimi K3 Model?

Kimi K3 is Moonshot AI’s flagship open-weight model, representing the first publicly available AI system in the 3-trillion-parameter class. The model utilizes a Mixture of Experts (MoE) architecture with 2.8 trillion total parameters, activating 104 billion parameters during inference through its Stable LatentMoE framework (16 of 896 experts per token).

The architecture incorporates two key innovations: Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals (AttnRes), both designed to improve information flow through longer sequences and deeper model layers. Moonshot claims these structural advances provide approximately 2.5x the scaling efficiency compared to K2.

Kimi K3 is a reasoning model that uses extended chain-of-thought thinking to work through complex problems. At launch, it operates at maximum thinking effort by default, with low- and high-effort modes planned for subsequent updates. The model supports multimodal input (text and images) and generates text output.

Technical Specifications

Specification	Value
Release Date	July 16, 2026
Developer	Moonshot AI
Total Parameters	2.8 trillion
Active Parameters	104 billion
Architecture	MoE with Stable LatentMoE
Expert Configuration	16 of 896 experts active
Context Window	1,048,576 tokens (1M)
Input Modalities	Text, Image
Output Modalities	Text
Output Speed	32.0 tokens/sec
Time to First Token	144.34s
License	Kimi K3 License (commercial use requires separate agreement)
Weights Availability	Open weights (released July 27, 2026)

API Provider Analysis

Kimi K3 is available through multiple API providers, each offering distinct advantages for different deployment scenarios. The following analysis compares the primary options available to developers.

Kimi (First-Party API)

Moonshot’s first-party API serves as the reference implementation for Kimi K3. All benchmark figures in this analysis derive from Kimi’s API performance.

• Input: $3.00 per 1M tokens
• Output: $15.00 per 1M tokens
• Cache Hit: $0.30 per 1M tokens (90% discount)
• Blended Rate: $2.31 per 1M tokens (7:2:1 cache/input/output ratio)
• Output Speed: 32.0 t/s
• TTFT: 222.66s
• Context: Full 1M context window support

The first-party API provides the official baseline behavior used in referenced benchmarks and is recommended when exact alignment with Moonshot’s specifications is required.

Fireworks

Fireworks AI launched Kimi K3 support on day zero of the open-weights release. The platform has developed custom optimizations specifically for K3’s architecture.

• Technical Optimizations: Custom kernels for Kimi Delta Attention, FP4 MoE kernels, adapted decode kernels for full K3 performance
• US-hosted infrastructure with zero data retention options
• Pay-per-token pricing (no dedicated GPU reservations required)
• LoRA adapter training with no setup required
• Model ID: accounts/fireworks/models/kimi-k3

Fireworks positions K3 as delivering performance comparable to closed frontier models at a meaningfully better cost-per-task. The platform is particularly suited for teams requiring immediate training capabilities alongside inference.

Together AI

Together AI provides reliable API endpoints optimized for open-weights models, with full support for K3’s extensive context window.

• Endpoint: moonshotai/Kimi-K3
• Input Price: $3.00/1M tokens
• Cache Hit: $0.30/1M tokens
• Output Price: $15.00/1M tokens
• Context Length: 1M tokens
• Quantization: FP4
• Capabilities: Preserved thinking history across turns, maximum thinking effort at launch, available on serverless and dedicated infrastructure

Together AI is recommended for applications requiring stable generation quality with preserved reasoning content across conversation turns.

OpenRouter

OpenRouter aggregates multiple Kimi K3 providers, offering flexible routing options for developers seeking optimal price-performance balance.

• Routing Modes: Balanced (optimizes for price and speed), Nitro (fastest response times), Exacto (highest tool-calling accuracy)
• Model ID: moonshotai/kimi-k3 (alias: ~moonshotai/kimi-latest)
• Full 1M-token context support, text and image input, tool calls, and structured outputs

OpenRouter is ideal for developers already using its API ecosystem who want automatic provider selection based on real-time performance metrics.

Provider Comparison Table

Provider	Best For	Key Advantage	Considerations
Kimi (First-Party)	Official baseline	Reference implementation	Higher pricing, slower performance
Fireworks	Training + Inference	Custom K3 kernels, US-hosted, ZDR	Day-zero support, LoRA training
Together AI	Long-horizon work	Preserved thinking history	Serverless and dedicated options
OpenRouter	Multi-provider routing	Automatic optimization	Aggregated pricing

Performance Benchmarks

Kimi K3 achieves a score of 57 on the Artificial Analysis Intelligence Index, a composite benchmark evaluating models across reasoning, knowledge, mathematics, and coding. This places K3 well above the median score of 32 for comparable open-weight models.

• Token usage: K3 generated 130M output tokens during Intelligence Index evaluation, well above the median of 63M (a reflection of its verbose, always-on chain-of-thought reasoning).
• Evaluation cost: $2,437.41 to run the full Intelligence Index.
• Comparative performance: Trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall performance, but outperforms Claude Opus 4.8 and GPT-5.5 on select benchmarks, and leads Arena.ai’s Code WebDev leaderboard.

Pricing Analysis

Kimi K3’s pricing positions it in the mid-to-premium tier for open-weight models:

Metric	Kimi K3	Open-Weight Median
Input (per 1M)	$3.00	$0.43
Output (per 1M)	$15.00	$1.20
Cache Hit (per 1M)	$0.30	—

• Cache economics: The 90% cache discount is significant for repetitive workloads.
• Flat context pricing: No long-context surcharge across the full 1M window.
• Effective blended rate: $2.31/M tokens with typical cache hit ratios.
• Versus Claude Opus: Approximately 40% cheaper on both input and output.
• Versus DeepSeek V4 Flash: Approximately 21x more expensive on output

Conclusion

Kimi K3 represents a compelling option for developers requiring frontier-level reasoning capabilities with the flexibility of open weights. Its massive 1M-token context window, multimodal input support, and strong benchmark performance make it particularly suited for complex, long-horizon tasks in coding, research, and enterprise workflows.

• For cost-sensitive deployments: Consider Fireworks for potentially lower effective pricing while maintaining access to K3’s capabilities.
• For training and fine-tuning: Fireworks offers the most streamlined path with LoRA adapter support and pay-per-token training.
• For production reliability: Together AI provides stable infrastructure with preserved thinking history, ideal for applications requiring consistent reasoning across sessions.
• For multi-model workflows: OpenRouter’s intelligent routing can optimize cost and performance across providers automatically.
• For official baseline behavior: Kimi’s first-party API remains the reference implementation for exact benchmark alignment.

The model’s primary limitations, high latency and slower output speed relative to peers, make it better suited for asynchronous, high-complexity tasks where reasoning accuracy is prioritized over immediate response times. For latency-sensitive applications, consider routing simpler queries to faster models while reserving K3 for complex reasoning tasks.

Kimi K3 is coming soon to DeepInfra. In the meantime, the previous generation, Kimi K2.5, is live on the platform, and the broader DeepInfra model catalog is worth monitoring for when K3 support goes live.