What are the strengths of this model?

Legendary high parameter count (over 2.5T parameters) Extended context length (over 1M tokens) High expectations as a next-gen flagship model

What are the weaknesses of this model?

Lack of official information Uncertain release date based on rumors Missing licensing and evaluation benchmarks

What are the best use cases?

General text generation Reasoning tasks Natural language understanding

Moonshot AIProprietary

Kimi K3

Name: Kimi K3
Author: Moonshot AI

Compare this model

Kimi K3 is Moonshot AI's open-weight flagship, released July 16, 2026 — the largest open-weight model to date at 2.8 trillion parameters (104B active per token across 896 experts) with a 1,000,000-token context window. Licensed under the modified MIT-style Kimi K3 License, it runs native multimodal input (text, image, video) with always-on thinking and leads the open-weights field at 57 on the Artificial Analysis Intelligence Index (rank 4 of 189). It tops frontend coding (Frontend Code Arena 1,679 Elo, #1 globally) and scores 88.3% on Terminal-Bench 2.1, 93.5% on GPQA Diamond, and 91.2% on BrowseComp. API pricing is $3 per million input (cache hit $0.30) and $15 per million output tokens.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-07-16

Japanese Language Capability

🌐Multilingual

General multilingual model. Basic Japanese processing is possible, but inferior to specialized models.

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

$15

Billing Mode: standard

Strengths

•Legendary high parameter count (over 2.5T parameters)
•Extended context length (over 1M tokens)
•High expectations as a next-gen flagship model

Weaknesses

•Lack of official information
•Uncertain release date based on rumors
•Missing licensing and evaluation benchmarks

Use Cases

•General text generation
•Reasoning tasks
•Natural language understanding

Deep Analysis

Arena Elo

1679

#1 in Frontend Code Arena

SWE-Bench Verified

~80%

Est. from coding benchmarks, vs Fable 5 ~87.6%

Input Price (cache-miss)

$3.00/1M

vs GPT-5.6 Sol: $5.00/1M

Output Price

$15.00/1M

vs Claude Fable 5: $50.00/1M

Parameters

2.8T

Largest open-weight model to date

Context Window

1M tokens

Flat pricing, no length tiers

Strengths

・World's largest open-weight model (2.8T params) with near-frontier performance.
・Exceptional long-horizon coding and agentic task performance (SWE Marathon, BrowseComp leader).
・Competitive pricing (~40-50% cheaper than GPT-5.6 Sol on output).

Weaknesses

・Independent third-party benchmarks still pending; some metrics lag top proprietary models.
・Higher hallucination rate (51%) reported in independent testing vs predecessor.
・Inference at scale requires substantial hardware; not a single-server deployment.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Fable 5	~1800+ (est.)	~87.6% (Opus 4.7 verified)	92.6%	$10/$50
GPT-5.6 Sol	~1750 (est.)	~80%	94.1%	$5/$30
Claude Opus 4.8	1600	~80%	91.0%	~$15/$75 (est.)

Overview

Kimi K3, released by Moonshot AI on July 16, 2026, is a landmark open-weight model and the first to cross the 2.8-trillion parameter threshold. Built on novel Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture with a 1-million-token context window, it targets frontier-level long-horizon coding, agentic workflows, and knowledge work. While Moonshot's own benchmarks show it trailing the very top proprietary models (Claude Fable 5, GPT-5.6 Sol), K3 consistently outperforms all other tested systems, including Claude Opus 4.8 and GPT-5.5. Its launch signals the functional closing of the capability gap between the best open-weight and closed-source models, offering enterprises a high-performance, self-hostable alternative at aggressive pricing ($3/$15 per million tokens). Kimi K3's strategic release, timed just before the 2026 World AI Conference, represents a major technical and business escalation. The architecture innovations deliver a reported 2.5x improvement in scaling efficiency over its predecessor, enabling strong performance in sustained multi-hour engineering tasks (leading SWE Marathon) and complex information retrieval (state-of-the-art on BrowseComp). The full open-weight release, expected by July 27, will allow the community to verify these claims and fine-tune the model for specific domains. However, the model's higher hallucination rate and the sheer infrastructure required for self-hosting present practical considerations for adopters. Pricing marks a significant shift from Kimi's earlier ultra-cheap models. K3 is priced at parity with Western mid-tier offerings like Claude Sonnet 5, undercutting top-tier models like GPT-5.6 Sol by 40-50%. This positions Kimi K3 as a value-oriented frontier model, especially for cost-sensitive, high-volume coding and research workloads where its flat-rate, million-token context window provides a distinct economic advantage over tiered-pricing competitors.

Benchmarks & Performance

Based on Moonshot AI's published benchmarks (max reasoning effort), Kimi K3 demonstrates frontier-level performance. It leads several key agentic and coding benchmarks but trails the top two proprietary models overall. | Benchmark Category | Benchmark | Kimi K3 (max) | Claude Fable 5 (max) | GPT-5.6 Sol (max) | Claude Opus 4.8 (max) | Notes | | :--- | :--- | :--- | :--- | :--- | :--- | :--- | | **Coding** | SWE Marathon | **42.0** | 35.0 | 39.0 | 40.0 | Measures sustained multi-hour coding. K3 leads. | | | Program Bench | **77.8** | 76.8 | 77.6 | 71.9 | K3 leads by a narrow margin. | | | Terminal Bench 2.1 | 88.3 | 84.6 | **88.8** | 84.6 | K3 trails GPT-5.6 Sol by half a point. | | | FrontierSWE | 81.2 | **86.6** | 71.3 | 66.7 | Fable 5 leads significantly. | | | DeepSWE | 67.5 | 70.0 | **73.0** | 59.0 | GPT-5.6 Sol leads. | | **Agentic** | BrowseComp | **91.2** | 88.0 | 90.4 | 84.3 | K3 achieves SOTA on this long-context retrieval test. | | | Automation Bench | **30.8** | 29.1 | 29.7 | 27.2 | K3 leads. | | | GDPval-AA v2 (Elo) | 1668 | **1760** | 1748 | 1600 | Fable 5 and Sol lead; K3 beats Opus 4.8. | | | AA-Briefcase (Elo) | 1548 | **1583** | 1495 | 1354 | K3 beats GPT-5.6 Sol and Opus 4.8. | | **Reasoning** | GPQA Diamond | 93.5 | 92.6 | **94.1** | 91.0 | K3 is competitive with the top models. | | **Vision** | OmniDocBench | **91.1** | 89.8 | 85.8 | 87.9 | K3 leads in document understanding. | Independent evaluation by Artificial Analysis places Kimi K3 at an Intelligence Index score of 57, on par with Claude Opus 4.8 and GPT-5.5, and fourth overall behind Fable 5 (60) and GPT-5.6 Sol (59). This largely corroborates Moonshot's positioning. K3's hallucination rate was noted to have increased to 51% in this testing.

Detailed Comparison

**Kimi K3 vs. Claude Fable 5:** * **Performance:** Fable 5 leads overall and on benchmarks like FrontierSWE and HLE-Full. It holds a verified SWE-Bench Verified ceiling (~87.6% on Opus 4.7). K3 competes closely on many coding and agentic tasks and leads on BrowseComp, SWE Marathon, and Automation Bench. * **Pricing:** K3 is dramatically cheaper: $3/$15 vs. Fable 5's ~$10/$50. Per-task costs are roughly half. * **Access:** K3 will be open-weight (weights expected July 27), allowing self-hosting and modification. Fable 5 is closed and API-only. * **Best For:** Choose K3 for cost-sensitive, high-volume tasks, long-context work, or open-weight needs. Choose Fable 5 for the highest verified performance on complex reasoning and multi-agent coding. **Kimi K3 vs. GPT-5.6 Sol:** * **Performance:** Sol leads on key reasoning (GPQA) and some agentic indices (GDPval-AA). K3 trades blows and wins on benchmarks like BrowseComp, SWE Marathon, and SpreadsheetBench 2. * **Pricing:** K3 is cheaper: $3/$15 input/output vs. Sol's $5/$30. Cached input is $0.30 vs. Sol's $0.50. * **Context:** Both offer ~1M token windows. K3's pricing is flat across context length, while Sol's may vary. * **Best For:** Choose K3 for better value on long-context coding and open-weight flexibility. Choose Sol for tasks requiring the highest agentic coding scores (e.g., Terminal Bench) or deep integration with the OpenAI ecosystem. **Kimi K3 vs. Claude Opus 4.8:** * **Performance:** K3 surpasses Opus 4.8 on Moonshot's reported benchmarks, including GDPval-AA, AA-Briefcase, and several coding tests. Independent testing places them in the same overall tier (~57 on Artificial Analysis Index). * **Pricing:** K3's pricing is comparable to or better than Sonnet 5, positioning it as a high-value alternative to Opus 4.8, which is typically priced much higher. * **Architecture:** K3 is vastly larger (2.8T vs. estimated 150-200B params) and open-weight. Opus 4.8 is closed and compact. * **Best For:** Choose K3 for frontier-level open-weight capabilities and cost efficiency. Choose Opus 4.8 for proven reliability in production multi-agent workflows and if its ecosystem integrations are critical.

Community Feedback

The developer and research community reaction is a mix of awe and cautious optimism. Key themes: 1. **A Paradigm Shift for Open-Source:** The consensus is that open-weight models have now entered frontier territory. One commentator noted, "Open source is no longer lagging six months behind Western closed-source models." 2. **Benchmark Scrutiny:** While Moonshot's numbers are impressive, the community is eagerly awaiting independent verification from Artificial Analysis, LMSYS, and other third-party labs. The use of different agent harnesses (KimiCode, Claude Code, Codex) in benchmarks has drawn some discussion. 3. **Pricing & Market Impact:** The end of "super cheap Chinese AI" is a noted trend. K3's pricing is seen as aggressive but reflective of its capabilities, putting direct pressure on OpenAI and Anthropic's pricing models. 4. **Technical Curiosity:** Engineers are particularly interested in the Kimi Delta Attention and Attention Residuals architectures, and the practical implications of deploying a 2.8T MoE model. Questions about inference throughput and real-world latency are common. 5. **Geopolitical Narrative:** The release is widely viewed as a significant milestone in China's AI development, showcasing algorithmic efficiency to overcome hardware constraints and positioning Moonshot as a global leader in open-source AI. *Notable Reactions:* Artificial Analysis confirmed K3's strong showing (57 on Intelligence Index) but flagged the increased hallucination rate. Early adopters have praised its coding and long-context abilities but note the need for careful prompt engineering.

Use Cases

1. **Large-Scale Codebase Refactoring & Monorepo Work:** K3 excels at long-horizon software engineering. Its 1M-token context can ingest entire repositories, and its leading performance on SWE Marathon (42.0) demonstrates sustained, multi-hour coding sessions with minimal oversight. It's ideal for navigating massive codebases, executing cross-file refactors, and orchestrating terminal tools over extended periods. Choose K3 over Claude Code or Codex for the largest projects where context window is a bottleneck. 2. **Autonomous Research & Complex Data Synthesis:** K3 can act as a research agent. It demonstrated the ability to reproduce complex astrophysics calculations by reading, cross-validating, and implementing pipelines from 20+ papers in hours. Use it for deep literature reviews, data synthesis from disparate sources, or generating interactive research reports with visualizations (like the 42-year ASIC industry analysis). Its BrowseComp SOTA score (91.2) is a key strength here. 3. **Cost-Efficient High-Volume Automation & Tooling:** For enterprise workloads requiring thousands of API calls—like document parsing, data extraction, or batch analysis—K3's pricing ($3/$15) makes it a compelling choice over more expensive models like GPT-5.6 Sol. Its strong performance on Automation Bench and SpreadsheetBench 2 supports this. The flat-rate, long-context pricing is particularly advantageous for processing lengthy documents. 4. **Prototyping Visual & Interactive Applications:** The "Vision in the Loop" capability allows K3 to iterate between code and live screenshots or video. It can generate full-stack applications, 3D games (like the Three.js open-world demo), or data visualizations by understanding visual output. Use it for rapid prototyping of interactive dashboards, frontend designs, or educational content where visual feedback is integral to the development process.

Latest News

**July 16, 2026:** Official launch of Kimi K3. Model is live on Kimi.com, Kimi Work, Kimi Code, and the Kimi API (identifier: `kimi-k3`). Max reasoning effort is the default; low/high-effort modes coming later. **July 16, 2026:** Pricing confirmed: $0.30/MTok (cache-hit input), $3.00/MTok (cache-miss input), $15.00/MTok (output). Automatic context caching is enabled, reporting >90% hit rate in coding workloads. **July 16, 2026:** Independent evaluation published by **Artificial Analysis**. Kimi K3 scores 57 on the Intelligence Index, placing 4th overall. Strong results on agentic benchmarks but noted increase in hallucination rate to 51%. **July 16, 2026:** Launch includes two agent variants: **K3 Max** for general chat and agent tasks, and **K3 Swarm Max** for large-scale parallel processing, extending the Agent Swarm concept from K2.6. **Upcoming (by July 27, 2026):** Full open-weight model release planned. Technical report detailing architecture, training, and evaluations to be published alongside weights. **Ecosystem:** Kimi Code CLI updated to v0.25.0 and v0.26.0 on launch day, adding expanded subagent tooling. A Chinese-market top-up promotion offering 10-30% bonus credits is active until August 12.

Kimi K3's strategic release, timed just before the 2026 World AI Conference, represents a major technical and business escalation. The architecture innovations deliver a reported 2.5x improvement in scaling efficiency over its predecessor, enabling strong performance in sustained multi-hour engineering tasks (leading SWE Marathon) and complex information retrieval (state-of-the-art on BrowseComp). The full open-weight release, expected by July 27, will allow the community to verify these claims and fine-tune the model for specific domains. However, the model's higher hallucination rate and the sheer infrastructure required for self-hosting present practical considerations for adopters.

Pricing marks a significant shift from Kimi's earlier ultra-cheap models. K3 is priced at parity with Western mid-tier offerings like Claude Sonnet 5, undercutting top-tier models like GPT-5.6 Sol by 40-50%. This positions Kimi K3 as a value-oriented frontier model, especially for cost-sensitive, high-volume coding and research workloads where its flat-rate, million-token context window provides a distinct economic advantage over tiered-pricing competitors.

Sources

Analysis generated: 2026-07-17