Kimi K3: The First Open 3T-Class Model — 2.8T Parameters, Frontier Coding, and $3/$15 Pricing

Executive Summary

On July 16, 2026, Moonshot AI announced Kimi K3, a model that fundamentally shifts the open-weight frontier. At 2.8 trillion parameters, K3 is the world's first open model to reach the 3T-class — surpassing the previous record held by DeepSeek V4-Pro (1.6T) and Qwen3-Max-Preview (~1T). More importantly, it delivers frontier-level performance: 67.5% on DeepSWE, 88.3% on Terminal-Bench 2.1, 81.2% on FrontierSWE, and 56% on Humanity's Last Exam — placing it within striking distance of Claude Fable 5 and GPT-5.6 Sol.

The model is built on two novel architectural innovations: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), combined with an extreme Mixture-of-Experts sparsity that activates only 16 of 896 experts via a Stable LatentMoE framework. This yields approximately 2.5× scaling efficiency compared to Kimi K2, allowing the model to convert compute into intelligence far more effectively than its parameter count would suggest.

K3 is available immediately through the Kimi API, Kimi.com, Kimi Work, and Kimi Code (a terminal-based coding agent). Pricing is $3.00 per million input tokens (cache-miss), $15.00 per million output tokens, and $0.30 per million for cache-hit input — with Moonshot reporting >90% cache hit rates in coding workloads thanks to Mooncake's disaggregated inference architecture. The full model weights will be released as open source by July 27, 2026, alongside a detailed technical report.

1. The 3T-Class Breakthrough

1.1 Why 2.8 Trillion Parameters Matters

Kimi K3 represents a step-function increase in open model scale. To put it in perspective:

Model	Total Parameters	Active Parameters	Architecture	Open Weights?
Kimi K3	~2.8T	~50B (16/896 experts)	MoE + KDA + AttnRes	July 27, 2026
DeepSeek V4-Pro	~1.6T	~25B	MoE	Yes (MIT)
Qwen3-Max-Preview	~1T	~4B	MoE	Yes
MiniMax M3	~428B	~23B	MoE	Yes
Gemini 3.5 Pro	Unknown (target 2M context)	Unknown	Unknown	No
Claude Fable 5	Unknown	Unknown	Unknown	No
GPT-5.6 Sol	Unknown	Unknown	Unknown	No

For nine of the past twelve months, Kimi models have set the upper bound of open-model sizes. K3 extends this lead dramatically — the gap between K3 (2.8T) and the next largest open model (DeepSeek V4-Pro at 1.6T) is 75% more parameters.

1.2 The Scaling Efficiency Gain

What makes K3's scale meaningful is not just the parameter count but the 2.5× improvement in scaling efficiency over Kimi K2. Moonshot reports that the combination of KDA, AttnRes, and extreme MoE sparsity allows the model to convert compute into intelligence more effectively than a naive scale-up would achieve.

2. Architecture: KDA, AttnRes, and Extreme MoE

2.1 Kimi Delta Attention (KDA)

KDA is the foundational attention mechanism in K3, designed to scale efficiently across sequence length. Unlike standard attention that computes all pairwise interactions, KDA provides a more efficient foundation for long-context processing — critical for the 1M token context window.

KDA poses new challenges for conventional prefix caching, which Moonshot has addressed with a custom implementation contributed to the vLLM community (to be released alongside the model weights).

2.2 Attention Residuals (AttnRes)

AttnRes selectively retrieves representations across model depth rather than accumulating them uniformly. This allows the model to:

• Access relevant information from any layer in the network
• Reduce information degradation in very deep models
• Improve reasoning quality across long sequences

Together with KDA, AttnRes forms the architectural backbone that enables K3 to scale beyond the trillion-parameter regime.

2.3 Stable LatentMoE with Extreme Sparsity

K3's MoE configuration is extreme: 896 experts total, with only 16 active per token (1.8% sparsity). At this level of sparsity, routing and optimization become first-order challenges. Moonshot addresses this with three innovations:

1. Quantile Balancing: Derives expert allocation directly from router-score quantiles, eliminating heuristic updates and a sensitive balancing hyperparameter.
2. Per-Head Muon: Extends Muon optimization by optimizing attention heads independently for more adaptive learning at scale.
3. Sigmoid Tanh Unit (SiTU) + Gated MLA: Improves activation control and attention selectivity.

2.4 Quantization-Aware Training

K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility. This enables deployment on a wider range of hardware while maintaining quality.

2.5 Training Infrastructure

Moonshot introduces a fully balanced expert-parallel training method with static shapes and no host synchronization on the critical path, preventing expert imbalance from degrading throughput at large expert-parallel scales. They recommend deploying K3 on supernode configurations with 64 or more accelerators for optimal inference efficiency.

3. Benchmark Performance

3.1 Coding Benchmarks

K3's coding performance places it in the upper tier of all models, open or closed:

Benchmark	Kimi K3	Fable 5	GPT-5.6 Sol	Opus 4.8	Grok 4.5	GLM-5.2
DeepSWE	67.5%	—	—	—	62.0%	—
Terminal-Bench 2.1	88.3%	84.3%	83.4%	78.9%	83.3%	—
ProgramBench	77.8%	—	—	—	—	—
FrontierSWE	81.2%	—	—	—	—	—
SWE Marathon	42.0%	—	—	—	29.0%	—
PostTrain Bench	36.6%	—	—	—	—	—

Key observations:

• Terminal-Bench 2.1: K3's 88.3% leads all compared models, including Fable 5 (84.3%) and GPT-5.6 Sol (83.4%).
• DeepSWE: At 67.5%, K3 exceeds Grok 4.5 (62.0%) and matches the level of the strongest proprietary models.
• SWE Marathon: At 42.0%, K3 significantly outperforms Grok 4.5 (29.0%), demonstrating strength on very long-running tasks.
• FrontierSWE: The 81.2% dominance score indicates K3 can handle the most difficult software engineering tasks.

3.2 Agentic Benchmarks

Benchmark	Kimi K3	Gemini 3.5 Flash	GPT-5.5	Claude Opus 4.7
MCP Atlas	84.2%	83.6%	75.3%	79.1%
BrowseComp	91.2%	—	—	—
Toolathlon-Verified	73.2%	—	—	—
AutomationBench	30.8%	—	—	—
JobBench	52.9%	—	—	—

On MCP Atlas (multi-step agentic workflows), K3's 84.2% slightly exceeds Gemini 3.5 Flash (83.6%) and significantly exceeds GPT-5.5 (75.3%) and Claude Opus 4.7 (79.1%).

3.3 Reasoning & Knowledge

Benchmark	Kimi K3	Fable 5	GPT-5.6 Sol	Opus 4.8	Sonnet 5
Humanity's Last Exam	56.0%	—	—	—	—
GPQA-Diamond	93.5%	—	—	—	—
HLE (no tools)	43.5%	—	—	—	—

The HLE score of 56% with tools is competitive with the strongest frontier models, indicating K3's reasoning capabilities are genuine, not just a product of tool use.

3.4 Independent Evaluations

Artificial Analysis independently evaluates K3:

• Intelligence Index: 57.1 (well above average for the price tier; median is 31)
• Coding Index: 76.2
• Agentic Index: 50.1
• GDPval-AA: 1685 ELO
• AA-Briefcase: 1548 ELO

BenchLM ranks K3 #4 out of 200 models on the public leaderboard with an overall score of 80.96/100, placing it behind only Claude Mythos 5 (83.93), Claude Fable 5 (83.68), and GPT-5.6 Sol (81.96).

3.5 Chatbot Arena

K3's LMSYS Chatbot Arena performance (as of July 19, 2026):

Category	Elo	Votes
Text Overall	1486	3,024
Coding	1529	798
Hard Prompts	1506	1,942
Hard Prompts (English)	1518	708
Longer Query	1498	1,308

The Coding Elo of 1529 places K3 among the top-rated models for coding tasks in the arena.

4. Pricing Analysis

4.1 K3 Pricing Structure

Variant	Cache-Hit Input	Cache-Miss Input	Output
Kimi K3	$0.30/M	$3.00/M	$15.00/M

Moonshot reports that the official Kimi API achieves a cache hit rate above 90% in coding workloads thanks to Mooncake's disaggregated inference architecture. This means effective input pricing in coding scenarios is approximately:

Effective input cost = (0.90 × $0.30) + (0.10 × $3.00) = $0.27 + $0.30 = $0.57/M

4.2 Competitive Pricing Context

Model	Input (per 1M)	Output (per 1M)	Cache Hit	Provider
DeepSeek V4-Flash	$0.14	$0.28	—	DeepSeek
MiniMax M2.7	$0.30	$1.20	—	MiniMax
Gemini 3.5 Flash	$1.50	$9.00	$0.15	Google
Grok 4.5	$2.00	$6.00	$0.50	xAI
Kimi K3 (effective)	~$0.57	$15.00	$0.30	Moonshot
Kimi K3 (worst case)	$3.00	$15.00	$0.30	Moonshot
Claude Sonnet 5	$2.00	$10.00	—	Anthropic
GPT-5.6 Terra	$2.50	$15.00	—	OpenAI
Claude Opus 4.8	$5.00	$25.00	—	Anthropic
GPT-5.6 Sol	$5.00	$30.00	—	OpenAI

4.3 The Cache Advantage

K3's pricing model is heavily dependent on cache hit rates. In coding workloads where the same context (codebase, instructions, tool definitions) is reused across many turns, the >90% cache hit rate dramatically reduces effective costs:

# Example: 200K token context with 90% cache hit over 10 turns
# Without caching: 200,000 × $3.00/M × 10 = $6.00
# With 90% cache hit: (20,000 × $3.00/M) + (180,000 × $0.30/M) = $0.06 + $0.054 = $0.114 per turn
# Total for 10 turns: $1.14 (81% savings vs. no caching)

4.4 Cost-Performance Analysis

K3's position is unique: near-frontier capability at mid-tier pricing with open weights coming. While the $15/M output price is higher than Grok 4.5 ($6/M) or Gemini 3.5 Flash ($9/M), the effective input cost (~$0.57/M with caching) is among the lowest for models in this capability tier.

For teams that can leverage the cache (coding workflows, multi-turn agents, long-context tasks), K3 offers exceptional value. For short, one-shot queries without context reuse, the $3/M input price is competitive but not dominant.

5. Real-World Capabilities: Case Studies

5.1 Kernel Optimization

K3 was tested on GPU kernel optimization tasks spanning AttnRes, KDA, and a 512-head-dimension MLA kernel across NVIDIA H200 and alternative GPGPU hardware. K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8, GPT-5.6 Sol, and GPT-5.5.

Notably, in the late stages of K3's own development, an early version of K3 handled the majority of the team's kernel optimization work — a form of self-improving development.

5.2 GPU Compiler Development

K3 developed MiniTriton, a compact Triton-like compiler from scratch, including:

• Its own tile-level IR layer over MLIR
• Optimization passes
• PTX code-generation pipeline

MiniTriton delivers performance on par with or better than Triton and torch.compile across supported roofline benchmarks, and sustains end-to-end nanoGPT training with stable convergence.

5.3 Chip Design

In a single 48-hour autonomous run, K3 designed a chip to serve a nano model built on its own architecture:

• Built, optimized, and verified using open-source EDA tools on the Nangate 45nm library
• Within 4 mm², the chip closes timing at 100 MHz
• Sustains over 8,700 tokens/s decode throughput in simulation
• Packs 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization

A chip built by a model, for a model.

5.4 Scientific Research

K3 completed in about two hours what would typically require one to two weeks of work by an experienced researcher:

• Reviewed and cross-validated 20+ papers on I–Love–Q universal relations in computational astrophysics
• Implemented the full numerical pipeline
• Evaluated 300+ equations of state
• Identified inconsistencies in published formulas
• Generated 3,000+ lines of Python code
• Produced an interactive HTML dashboard

5.5 Interactive Research Reports

K3 created an interactive research report on 42 years of the AI ASIC industry through 120+ rounds of recursive self-improvement:

• 2,800+ web searches/fetches
• 1,100+ terminal data pulls
• 11,000+ pages across 87 quarterly reports and 99 original PDFs
• Bespoke charts, animated diagrams, and interactive visual narratives

6. Deployment Guidance

6.1 API Access

K3 is available through the OpenAI-compatible Kimi API:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

completion = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "user", "content": "Debug this production issue..."}
    ],
    reasoning_effort="max",  # Currently the only supported value
)

print(completion.choices[0].message.content)

6.2 Platform Availability

Platform	Availability	Notes
Kimi API	Available	OpenAI-compatible, $3/$15 pricing
Kimi.com	Available	Web interface
Kimi Work	Available	Desktop app (Windows + Mac), v3.1.0+
Kimi Code	Available	Terminal-based coding agent
Kimi App	Available	iOS, Android, HarmonyOS
Kimi Enterprise	Available	Data privacy, member management
OpenRouter	Available	Via model gateway
Hugging Face	Coming July 27	Full open weights

6.3 Parameter Constraints

K3 has fixed inference parameters — they cannot be modified:

Parameter	Value	Modifiable?
temperature	1.0	No
top_p	0.95	No
n	1	No
presence_penalty	0	No
frequency_penalty	0	No
reasoning_effort	"max"	No (only value currently)
tool_choice	auto/none/required	Yes

6.4 Multimodal Support

K3 supports text, image, and video input natively:

{
  "model": "kimi-k3",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "image_url",
          "image_url": {"url": "data:image/png;base64,..."}
        },
        {
          "type": "text",
          "text": "Please describe this image and fix the bug shown."
        }
      ]
    }
  ]
}

6.5 Context Management

K3 supports a 1M token context window. For very long conversations, Moonshot recommends context compaction triggered at 300K tokens (as demonstrated in their BrowseComp evaluation, where the 1M-context run without context management scored 90.4% vs. 91.2% with compaction).

7. Integration with Prior Research

The Kimi K3 launch connects to several ongoing research threads:

• Gemini 3 5 Flash Frontier Agents Coding Flash Tier Cost 2026 07 17: Gemini 3.5 Flash leads on MCP Atlas (83.6% vs. K3's 84.2% — K3 actually edges ahead). Flash's $1.50/$9 pricing is more aggressive, but K3's 2.8T open-weight architecture and 1M context with native vision provide capabilities Flash doesn't match. The open weights release on July 27 will be a game-changer for self-hosting.

• Minimax M27 Self Evolving Agent Harness Open Weight Frontier 2026 07 16: M2.7's $0.30/$1.20 pricing and self-evolution methodology are impressive, but K3's 2.8T scale and 67.5% DeepSWE score represent a significant capability jump. M2.7 is the cost champion; K3 is the capability champion among open models.

• Grok 4 5 Cursor Trained Moe Coding Agentic Knowledge Work 2026 07 15: Grok 4.5's 62.0% DeepSWE and 83.3% Terminal-Bench are solid, but K3 exceeds both (67.5% and 88.3%). Grok's 4.2× token efficiency advantage may still make it more cost-effective per task, but K3's raw capability is higher.

• Claude Sonnet 5 Most Agentic Sonnet 1m Context Adaptive Thinking 2026 07 14: Sonnet 5's 85.2% SWE-bench Verified and adaptive thinking are strong, but K3's 1M context, native vision, and open-weight trajectory create a different value proposition. Sonnet 5 at $2/$10 is cheaper on output, but K3's caching brings effective costs down.

• Deepseek V4 Flash Pro Api Migration July 24 Deadline Architecture Pricing 2026 07 10: DeepSeek V4-Pro at 1.6T parameters is now dwarfed by K3's 2.8T. The July 24 migration deadline for DeepSeek's legacy models creates an interesting comparison: V4-Pro ($0.435/$0.87) is cheaper but K3's capability is significantly higher.

• Gemini 3 5 Pro Rebuilt Frontier 2m Context Deep Think July 17 Showdown 2026 07 13: Gemini 3.5 Pro's continued delays contrast sharply with K3's successful launch. K3's 1M context is smaller than Pro's target 2M, but K3 has shipped and is in production.

8. Key Takeaways

1. The open-weight frontier has jumped to 3T: K3's 2.8T parameters represent a 75% increase over the previous largest open model (DeepSeek V4-Pro at 1.6T). This gap may widen further as K3's open weights are released on July 27.

2. Frontier-level coding is now open: K3's 67.5% DeepSWE, 88.3% Terminal-Bench, and 81.2% FrontierSWE scores place it within striking distance of Fable 5 and GPT-5.6 Sol — and these are from a model that will be freely downloadable.

3. The cache pricing model is a game-changer: With >90% cache hit rates in coding workloads, K3's effective input cost drops to ~$0.57/M — making it one of the most cost-effective frontier models for multi-turn coding and agentic workflows.

4. KDA and AttnRes are architectural breakthroughs: These novel attention mechanisms enable K3 to scale efficiently beyond the trillion-parameter regime. The 2.5× scaling efficiency gain over K2 suggests these innovations have broad applicability.

5. Real-world case studies are impressive: From GPU kernel optimization to chip design to scientific research, K3's capabilities extend far beyond benchmark scores. The 48-hour autonomous chip design and 2-hour astrophysics pipeline are particularly noteworthy.

6. The open weights release on July 27 is critical: Once K3's weights are available, the self-hosting community will have access to the most capable open model ever released. This could accelerate the open-weight frontier significantly.

7. Moonshot's momentum is real: For nine of the past twelve months, Kimi models have set the upper bound of open-model sizes. K3 extends this lead dramatically, and the promised technical report will provide deeper insights into the architecture.

9. Future Directions

9.1 Immediate (July 20-27, 2026)

• Open weights release (July 27): The full model weights and technical report will be published. This is the most anticipated event in the open-weight community this month.
• Independent benchmark replication: The vendor-reported scores need independent verification. Watch for evaluations from yage.ai, BenchLM, and Scale AI's SEAL leaderboard.
• GGUF quantization: Once weights are released, expect rapid GGUF quantizations from Unsloth, llama.cpp, and Ollama communities.
• Qwen 3.8 comparison: Alibaba's Qwen 3.8 (2.4T params) was announced July 19 with no benchmarks yet. A direct comparison between K3 and Qwen 3.8 will be the defining open-model showdown of Q3 2026.

9.2 Medium-Term (Q3 2026)

• Reasoning effort levels: Moonshot has promised low- and high-effort modes in subsequent updates. This will enable more granular cost control.
• Enterprise adoption: Will K3's combination of frontier capability and open weights drive enterprise migration from more expensive closed models?
• vLLM integration: The KDA-compatible prefix caching implementation will be released alongside the model, enabling efficient serving.
• Multi-agent ecosystem: K3's strong agentic benchmarks (84.2% MCP Atlas, 91.2% BrowseComp) could spark new multi-agent frameworks built on open-weight models.

9.3 The Bigger Picture

K3 represents a strategic shift in the open-weight frontier: from "good enough at low cost" to "frontier-level at accessible cost". The combination of 2.8T parameters, novel architecture (KDA + AttnRes), extreme MoE sparsity, and open weights creates a scenario where the gap between open and closed models continues to narrow.

If the open weights release on July 27 goes smoothly, K3 could become the default choice for teams that need frontier-level coding and agentic capabilities without the vendor lock-in and premium pricing of closed models. The self-hosting community's ability to fine-tune, optimize, and deploy K3 on their own infrastructure could accelerate innovation in ways that closed models cannot match.

References & Resources

Official Sources

1. Moonshot AI. (2026). Kimi K3 Tech Blog: Open Frontier Intelligence. https://www.kimi.com/blog/kimi-k3
2. Moonshot AI. (2026). Kimi API Platform. https://platform.kimi.ai/
3. Moonshot AI. (2026). Kimi API Documentation — Models Overview. https://platform.kimi.ai/docs/api/models-overview
4. Moonshot AI. (2026). Kimi API Quickstart. https://platform.kimi.ai/docs/overview.md
5. Moonshot AI. (2026). Kimi.com. https://www.kimi.com/
6. Moonshot AI. (2026). Kimi Work. https://www.kimi.com/products/kimi-work
7. Moonshot AI. (2026). Kimi Code. https://www.kimi.com/code
8. Moonshot AI. (2026). Kimi Enterprise. https://www.kimi.com/membership/pricing

Independent Evaluations

9. BenchLM. (2026). Kimi K3 Benchmarks, Pricing & Speed. https://benchlm.ai/models/kimi-3
10. Artificial Analysis. (2026). Kimi K3 — Intelligence, Performance & Price Analysis. https://artificialanalysis.ai/models/kimi-k3
11. OpenRouter. (2026). Kimi K3 — API Pricing & Benchmarks. https://openrouter.ai/moonshotai/kimi-k3

Related Research

• Gemini 3 5 Flash Frontier Agents Coding Flash Tier Cost 2026 07 17
• Minimax M27 Self Evolving Agent Harness Open Weight Frontier 2026 07 16
• Grok 4 5 Cursor Trained Moe Coding Agentic Knowledge Work 2026 07 15
• Claude Sonnet 5 Most Agentic Sonnet 1m Context Adaptive Thinking 2026 07 14
• Deepseek V4 Flash Pro Api Migration July 24 Deadline Architecture Pricing 2026 07 10
• Gemini 3 5 Pro Rebuilt Frontier 2m Context Deep Think July 17 Showdown 2026 07 13
• Ai News Week 2026 07 06 2026 07 13

---

This article was researched and written on July 20, 2026, based on the official Kimi K3 tech blog post (https://www.kimi.com/blog/kimi-k3), Kimi API documentation (https://platform.kimi.ai/docs/), and independent evaluations from BenchLM.ai and Artificial Analysis. All benchmark figures are sourced from Moonshot AI's official launch announcement and cross-referenced with independent evaluation platforms.