# Kimi K3: Moonshot AI's 2.8T Open Frontier Intelligence Model

Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8 trillion parameter MoE model with 1M token context, native vision, and frontier-level benchmarks. Full analysis, benchmarks, and pricing breakdown.

On July 16, 2026, Moonshot AI released **Kimi K3** — a 2.8 trillion parameter mixture-of-experts model that the company calls "the world's first open 3T-class model." Available immediately via API and the Kimi app, with full open weights promised by July 27, 2026, K3 represents Moonshot's most aggressive step yet at the scaling frontier.

According to Moonshot's official benchmark table, K3 consistently outperforms other tested models including Claude Opus 4.8 and GPT 5.5, while remaining behind the two current leaders — Claude Fable 5 and GPT 5.6 Sol. Independent analysis from Artificial Analysis places K3 at **#4 out of 189 models** on their Intelligence Index.

2.8T

Total Parameters

1M

Context Window

16/896

Experts Active

$3/$15

Per 1M Tokens

## Architecture: Kimi Delta Attention and Scaling Efficiency

Kimi K3 is built on two key architectural innovations: **Kimi Delta Attention (KDA)** and **Attention Residuals (AttnRes)**. KDA provides an efficient foundation for scaling attention across sequence length, while AttnRes selectively retrieves representations across model depth rather than accumulating them uniformly.

The model uses a **Stable LatentMoE** framework, activating 16 of 896 experts per token. At this level of sparsity (less than 2%), routing and optimization become first-order challenges. Moonshot addresses this with:

- **Quantile Balancing** — derives expert allocation directly from router-score quantiles, eliminating heuristic updates and a sensitive balancing hyperparameter

- **Per-Head Muon** — extends Muon optimizer by optimizing attention heads independently for more adaptive learning at scale

- **Sigmoid Tanh Unit (SiTU)** — improves activation control

- **Gated MLA** — enhances attention selectivity

Together, these structural changes yield an approximate **2.5× improvement in overall scaling efficiency** compared to Kimi K2, according to Moonshot. The model applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility. Source: Moonshot AI official blog

## Benchmark Performance

Moonshot published an extensive benchmark table comparing K3 against Claude Fable 5, GPT 5.6 Sol, Claude Opus 4.8, GPT 5.5, and GLM-5.2. All Kimi K3 results use **max reasoning effort**, temperature 1.0, top-p 1.0.

### Reasoning & Knowledge

| **Benchmark** | **K3 (max)** | **Fable 5** | **GPT 5.6 Sol** | **Opus 4.8** | **GPT 5.5**|
--- | --- | --- | --- | --- | ---
| GPQA-Diamond | 93.5 | 92.6 | 94.1 | 91.0 | 93.5|
| HLE-Full | 43.5 | 53.3 | 44.5 | 49.8* | 41.4*|
| HLE-Full w/ tools | 56.0 | 63.0 | 58.0 | 57.9* | 52.2*|

### Coding

| **Benchmark** | **K3 (max)** | **Fable 5** | **GPT 5.6 Sol** | **Opus 4.8** | **GPT 5.5**|
--- | --- | --- | --- | --- | ---
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | 67.0|
| Terminal-Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 | 83.4|
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 64.9|
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 14.0|
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 | 70.8|

### Agentic & Knowledge Work

| **Benchmark** | **K3 (max)** | **Fable 5** | **GPT 5.6 Sol** | **Opus 4.8** | **GPT 5.5**|
--- | --- | --- | --- | --- | ---
| GDPval-AA v2 (Elo) | 1668 | 1760 | 1748 | 1600 | 1494|
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | 84.4|
| Toolathlon-Verified | 73.2 | 77.9 | 74.9 | 76.2 | 73.5|
| MCP Atlas | 84.2 | 84.7 | 83.6 | 83.6 | 82.8|
| AA-Briefcase (Elo) | 1548 | 1583 | 1495 | 1354 | 1158|

### Multimodal & Vision

| **Benchmark** | **K3 (max)** | **Fable 5** | **GPT 5.6 Sol** | **Opus 4.8** | **GPT 5.5**|
--- | --- | --- | --- | --- | ---
| MMMU-Pro | 81.6 | 81.2 | 83.0 | 78.9 | 81.2|
| CharXiv (RQ) | 84.8 | 88.9 | 84.6 | 80.5 | 84.1|
| MathVision | 94.3 | 94.8 | 95.8 | 86.7 | 92.2|
| WorldVQA (ForceAnswer) | 51.0 | 56.7 | 41.8 | 39.1 | 38.5|

*Source: Moonshot AI official blog (kimi.com/blog/kimi-k3). All scores self-reported by Moonshot. * indicates results cited from official model cards or third-party sources per Moonshot's footnotes.* Self-reported benchmarks

#### Key Takeaways

Kimi K3 **beats Claude Opus 4.8** on most benchmarks, establishing itself as a clear #3 in the frontier tier. Its **BrowseComp score of 91.2%** is the best published on this benchmark. On Arena.ai's Frontend Code arena, K3 has surpassed even Claude Fable 5.

However, **Claude Fable 5 and GPT 5.6 Sol** still hold the top positions across the benchmark suite. Moonshot itself acknowledges this gap in their limitations section.

## Independent Analysis: Artificial Analysis

Artificial Analysis evaluated Kimi K3 on their Intelligence Index v4.1 (which includes GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, HLE, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR). Their findings:

#4

of 189 Models

57

Intelligence Index

62.0

Tokens/Second

$2,691

Cost per Full Eval

Key observations from Artificial Analysis:

- **Intelligence: 57/100** — well above average (median: 30) among comparable reasoning models in the same price tier

- **Speed: 62.0 t/s** — below average (median: 72.7 t/s) for models in this price range

- **Verbosity: very high** — generated 130M output tokens on the Intelligence Index vs. 63M average, suggesting the model tends toward verbose reasoning

- **TTFT: 1.99s** — better than average (median: 2.60s) for its price tier

- **Cost per task: $2,690.80** — expensive due to verbosity, though cache hits bring effective input price to $0.30/MTok

Simon Willison's hands-on testing confirmed the verbosity issue: his pelican SVG prompt consumed 13,241 reasoning tokens to produce 3,417 output tokens, costing $0.25 for a task that would cost pennies on less verbose models. He also noted that the prompt "hi" counts as 86 tokens on K3 (vs. 10 for OpenAI and Anthropic models), suggesting an 85-token hidden system prompt that K3 refused to leak. Independent verification

## Pricing Analysis

#### Kimi K3 API Pricing (per 1M tokens)

$0.30

Cache Hit Input

$3.00

Cache Miss Input

$15.00

Output

At $3/$15 per million tokens, K3 is priced at the same level as Anthropic's Claude Sonnet series — making it the **most expensive model released by a Chinese AI lab to date**. This represents a significant increase over Kimi K2.6 ($0.95/$4) and K2.7 Code ($0.95/$4.00).

However, Moonshot reports that their API achieves a **cache hit rate above 90%** in coding workloads, powered by Mooncake's disaggregated inference architecture. With KDA-compatible prefill caching contributed to vLLM, effective costs can be dramatically lower for repeated codebases and documentation.

Artificial Analysis estimates a **blended rate of $2.31 per 1M tokens** using a 7:2:1 cache hit/input/output ratio — competitive with mid-tier models despite the list price.

#### Cost Consideration

K3 currently only supports **max thinking effort** — there is no low or medium mode. This means every response includes extended reasoning, which significantly increases token consumption. Simon Willison's testing showed the model using 13,241 reasoning tokens for a simple SVG generation task. Users should expect higher per-task costs compared to models with tiered reasoning options.

## Two Variants: K3 Max and K3 Swarm Max

The Kimi app offers two K3 variants:

- **K3 Max** — "Chat & Agent, flagship all-rounder." The default variant for everyday coding, reasoning, and assistant work.

- **K3 Swarm Max** — "Massive search, batch processing, and more in one go." The heavy variant for complex multi-agent workflows, fanning one task across many sub-agents in parallel.

Agent Swarm is K3's headline capability — coordinating many sub-agents to run search and processing in parallel. It extends the swarm system Moonshot shipped with Kimi Work, which already ran hundreds of coordinated sub-agents.

## Context Window: A Membership Tier

One of K3's most distinctive features is that **context length is gated by membership tier**:

- **Moderato** plan: 256K token context

- **Allegretto and higher** plans: 1M token context

- Full API access: 1M token context (no tier restriction)

This is a notable shift in how AI pricing is evolving: the model is the CPU, the context window is the RAM, and your membership tier is the RAM configuration. A 1M context window represents roughly **1,573 A4 pages of 12pt Arial font** — enough to hold an entire codebase, its documentation, logs, and task history in working memory at once.

## Notable Demonstrations

Moonshot published several impressive demonstrations of K3's capabilities:

### GPU Compiler Development (MiniTriton)

K3 built MiniTriton — a compact Triton-like compiler with its own tile-level IR layer over MLIR, optimization passes, and a PTX code-generation pipeline. Across supported roofline benchmarks, MiniTriton delivers performance on par with or better than Triton and torch.compile. It sustained end-to-end nanoGPT training with stable convergence. Self-reported

### Chip Design

In a 48-hour autonomous run, K3 designed a chip to serve a nano model built on its own architecture. Using open-source EDA tools on the Nangate 45nm library, the chip closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation — 1.46M standard cells, 0.277 MB SRAM, and an INT4 MAC array. A chip built by a model, for a model. Self-reported

### Research Automation

K3 reproduced the I–Love–Q universal relations in computational astrophysics in approximately 2 hours — work that typically requires 1-2 weeks of an experienced researcher's time. It reviewed 20+ papers, implemented the full numerical pipeline, evaluated 300+ equations of state, and generated 3,000+ lines of Python code. Self-reported

### Game Development

K3 demonstrated the ability to build fully playable browser-based 3D games, GBA emulators, and cyberpunk web-swinging experiences using Three.js WebGPU and GPU compute — combining 3D reasoning, coding, and vision capabilities in a "vision in the loop" workflow.

## Limitations and Caveats

#### Important Limitations (from Moonshot's official blog)

**Sensitivity to thinking history:** K3 was trained in preserved thinking history mode. If an agent harness fails to pass back all historical thinking content, or if a session switches from another model to K3 mid-conversation, generation quality may become highly unstable. Moonshot recommends using a compatible harness (Kimi Code) and avoiding mid-session switches.

**Excessive proactiveness:** K3's training emphasizes long-horizon challenging tasks, which means it may make unexpected decisions on the user's behalf when encountering minor issues or ambiguous intent. Applications requiring well-defined boundaries should impose explicit behavioral constraints in system prompts or AGENTS.md.

**Gap with top models:** Despite being highly competitive overall, K3 "nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol."

Additional practical limitations observed by independent testers:

- **Only max thinking effort available** — no low or medium mode, leading to high token consumption and longer response times

- **Below-average speed** — 62 t/s vs. 72.7 median for models in its price tier (Artificial Analysis)

- **Very verbose output** — 130M tokens on Intelligence Index vs. 63M average

- **No official model card yet** — full architectural and training details promised alongside the open weights release on July 27

## Availability

Kimi K3 is available through multiple channels:

- **Kimi app** (iOS, Android, HarmonyOS) — K3 Max and K3 Swarm Max

- **Kimi Work** (Windows, Apple silicon Mac) — desktop agent for knowledge work, v3.1.0+

- **Kimi Code** — CLI and IDE agent, select with `/model` command

- **Kimi API** (`kimi-k3`) — OpenAI-compatible endpoint at platform.kimi.ai

- **OpenRouter** — available via `moonshotai/kimi-k3`

- **Open weights** — promised by July 27, 2026 (not yet released)

## Conclusion

Kimi K3 represents a significant milestone in the open frontier race. At 2.8 trillion parameters with native vision and a 1M token context window, it delivers frontier-level performance that consistently beats Claude Opus 4.8 and approaches the capabilities of Claude Fable 5 and GPT 5.6 Sol.

Its strengths are most pronounced in **coding** (Terminal-Bench 88.3%, SWE Marathon 42%), **agentic knowledge work** (BrowseComp 91.2%, GDPval-AA 1668), and **long-horizon tasks** where its massive context window provides a real advantage. The Agent Swarm architecture for parallel multi-agent workflows is a unique differentiator.

The trade-offs are clear: **higher cost** (especially with max-only thinking), **below-average speed**, **verbose output**, and a **current gap with the two top models**. The open weights release on July 27, 2026 will be the true test of K3's practical value — enabling self-hosting, fine-tuning, and ecosystem integration that proprietary models can't match.

For users who need frontier reasoning at a lower price than Opus or Fable, and who value the open-weight path, K3 is worth serious consideration. For those who need the absolute best performance regardless of cost, Fable 5 and GPT 5.6 Sol still lead.

#### Our Assessment

**Best for:** Long-horizon coding, agentic workflows, users who value open-weight flexibility, cost-conscious users who need near-frontier performance.

**Not ideal for:** Quick chat tasks (max-only thinking), latency-sensitive applications (below-average speed), users who need tiered reasoning effort.

**Watch for:** Open weights release (July 27), incoming low/medium thinking modes, community benchmarks from independent evaluators.

### Related Coverage

- [Kimi K2.7 Code: Moonshot's Dedicated Coding Model](/kimi-k2-7-code-moonshots-new-ai-model-built-for-coding.html)

- [Claude Fable 5: Anthropic's New Flagship](/claude-fable-5-suspended-us-government-export-control.html)

- [GPT 5.6 Sol: OpenAI's Latest Release](/gpt-5-6-openai-sol-terra-luna.html)

By ZVHH Research — AI Tools & Software Reviews

Benchmarks sourced from Moonshot AI official blog, Artificial Analysis, OpenRouter, and independent community testing.

Powered by Allam 7B (SDAIA)