Moonshot AI has just dropped Kimi K3, and it’s a monster. At 2.8 trillion parameters, it’s the world’s first open-source model to break into the 3-trillion-parameter class. Released July 16, 2026, Kimi K3 pushes the boundaries of what open-weight models can do in coding, knowledge work, and long-horizon agentic workflows. Full open-weight release follows by July 27.
The Scale: 2.8 Trillion Parameters
Let that number sink in. Kimi K3 has 2.8 trillion total parameters, making it the largest open-source model available. For context, the previous Kimi K2 series sat at around 1 trillion. Moonshot AI has been on a relentless scale push — for nine of the past twelve months, Kimi models have set the upper bound of open-model sizes.
But raw scale alone doesn’t tell the story. The architectural innovations are what make 2.8T parameters actually usable.
Architecture: KDA, Attention Residuals, and Stable LatentMoE
Kimi K3 introduces three key architectural innovations that together deliver roughly 2.5x the overall scaling efficiency of K2:
- Kimi Delta Attention (KDA) — A hybrid linear attention mechanism that helps information flow more smoothly through longer sequences and deeper models. Moonshot has open-sourced high-performance KDA kernels via FlashKDA, built on CUTLASS for SM90+ GPUs. This is the core efficiency enabler that makes the massive parameter count tractable.
- Attention Residuals (AttnRes) — A complementary technique that selectively retrieves representations across depth rather than accumulating them uniformly, reducing the degradation that typically plagues very large models.
- Stable LatentMoE — An extreme sparsity framework where only 16 out of 896 experts are activated per token. This means that despite the 2.8T total parameter count, the active parameter footprint per forward pass is dramatically smaller, keeping inference costs reasonable.
The training stack also includes Quantile Balancing for expert allocation (deriving allocation from router-score quantiles), Per-Head Muon for independent attention-head optimization, Sigmoid Tanh Unit (SiTU) for activation control, and Gated MLA for attention selectivity. The model applies quantization-aware training from the SFT stage onward using MXFP4 weights with MXFP8 activations.
Benchmarks: Top-Tier Across the Board
According to Artificial Analysis benchmarks aggregated on OpenRouter, Kimi K3 lands in the top percentiles across multiple categories:
| Benchmark | Score | Percentile |
|---|---|---|
| Intelligence Index | 57.1 | Top 3% of all models |
| Coding Index | 76.2 | Top 5% of all models |
| Agentic Index | 50.1 | Top 3% of all models |
| GPQA Diamond (graduate-level science) | 93.5% | — |
| HLE (Humanity’s Last Exam) | 44.3% | — |
| AA-LCR (long context reasoning) | 74.7% | — |
| GDPval-AA (economically valuable tasks) | 59.2% | — |
| SciCode (scientific Python) | 58.7% | — |
All benchmark results were obtained with reasoning effort set to max. The 93.5% on GPQA Diamond is particularly notable — that’s graduate-level scientific reasoning, and it puts Kimi K3 in elite territory alongside closed models from OpenAI and Anthropic. The Coding Index of 76.2 confirms this is a serious coding model, not just a chatbot that happens to write code.
1M Token Context with Automatic Caching
Kimi K3 supports a 1,048,576-token context window — the full million. This is critical for the agentic workflows the model is designed for: navigating large code repositories, processing extensive documentation, and maintaining context across multi-step tool interactions.
Context caching is automatic. No cache IDs, no TTLs, no extra parameters. Just keep the long prefix unchanged across requests, and Moonshot’s infrastructure handles the rest. On OpenRouter, this is already delivering a 92.5% cache hit rate, bringing the weighted average effective input price down to $0.502 per million tokens. Moonshot’s own official API reports a cache hit rate above 90% in coding workloads.
Pricing and Performance
| Metric | Value |
|---|---|
| Input price | $3.00 / 1M tokens |
| Output price | $15.00 / 1M tokens |
| Cache read | $0.30 / 1M tokens |
| Effective input (weighted avg, with caching) | ~$0.50 / 1M tokens |
| Throughput | 12 tokens/sec |
| Latency (p50) | 16.47 seconds |
| Provider uptime (30d) | 99.92% |
At $3/$15 per million tokens, Kimi K3 is priced competitively for a frontier-class model. The aggressive caching discount — over 90% of input tokens hit cache in coding workloads — makes it even more attractive for agentic use cases where the system prompt and tool definitions are repeated across calls. Performance figures are from OpenRouter’s live metrics as of July 20, 2026.
Native Multimodal and Agentic Capabilities
Kimi K3 has native visual understanding. It accepts images via base64 encoding (public URLs are not supported for vision input) and can process video files. The model is particularly strong at tasks that combine software engineering with visual reasoning — using screenshots and visual feedback to improve workflows in game development, frontend engineering, and CAD.
The API is fully OpenAI-compatible and available through the Kimi API Platform. Here’s a basic call:
from openai import OpenAI
client = OpenAI(
api_key="your-moonshot-api-key",
base_url="https://api.moonshot.ai/v1",
)
completion = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Explain quantum entanglement."}],
reasoning_effort="max",
)
print(completion.choices[0].message.content)Thinking Mode and Reasoning Effort
Kimi K3 always has thinking mode enabled. Unlike K2.x where thinking was optional, K3 reasons on every request. At launch, the default and only supported reasoning effort level is max, with low- and high-effort modes promised in subsequent updates. The thinking process is exposed via a separate reasoning_content field in streaming responses, so you can observe the chain-of-thought in real time.
Important constraint: temperature, top_p, n, presence_penalty, and frequency_penalty are all fixed and should be omitted from requests. This is a deliberate design choice — the model is tuned for its default sampling parameters.
Developer Experience
Moonshot has clearly invested in developer ergonomics. The model supports:
- Tool calls with
tool_choicecontrol (use"required"to force tool use on the first turn) - Dynamic tool loading — inject tool definitions into system messages without pre-registering them
- Structured output with strict JSON Schema enforcement
- Partial mode — continue generation from a text prefix
- Official tools via the Kimi API Platform (web search, code runner, memory, and more)
- Streaming with separate
reasoning_contentand answer deltas
Integration guides are available through the Kimi API Platform documentation for popular coding agents including Kimi Code CLI, Claude Code, Codex, OpenCode, and Hermes Agent. The model is also accessible through OpenRouter at moonshotai/kimi-k3.
The Bigger Picture
Kimi K3 represents a significant milestone for open-source AI. For nine of the past twelve months, Kimi has held the scale frontier among open models. The architectural innovations — KDA, AttnRes, Stable LatentMoE with 16/896 expert activation — show that scaling is not just about throwing more parameters at the problem. It’s about making those parameters efficient.
The 2.5x scaling efficiency improvement over K2 is the number to watch. If Moonshot can sustain that trajectory, we’re looking at a model family that converts compute into capability more effectively than any other open-source line. The full model weights arrive by July 27, 2026, along with a technical report detailing the architecture, training methodology, and evaluation.
For developers building agentic systems, coding assistants, or knowledge-work pipelines, Kimi K3 is now a serious contender. The combination of 1M context, native multimodal, always-on reasoning, and aggressive caching pricing makes it particularly well-suited for production workloads where context windows are large and repeated.
The open-source frontier just moved. Again.