Kimi K3: 2.8T Params, Open-Source, Taking on GPT-5.6

On July 16, 2026, Beijing-based Moonshot AI dropped Kimi K3 — a 2.8-trillion-parameter model that instantly became the largest open-source model ever released. The announcement, timed just ahead of the 2026 World AI Conference in Shanghai, positions K3 as a direct challenger to Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. And unlike many “open” releases that stop at a research paper, Moonshot is shipping full model weights by July 27.

TL;DR: Kimi K3 is 75% larger than DeepSeek V4 Pro (2.8T vs 1.6T params), beats Claude Opus 4.8 across most benchmarks, and costs $3/M input tokens — matching Anthropic’s Sonnet pricing. It ranked #1 on BrowseComp, #1 on Frontend Code Arena, and used its own early version to optimize GPU kernels during training. Open weights drop July 27. This is the biggest open-source AI release of 2026.

Kimi K3 Scale: 2.8T Parameters vs the Competition

Let’s put this in perspective. Here’s how K3 stacks up against other open-weight models:

Model	Parameters	Context	Release Date	Pricing (Input/Output per 1M tokens)
Kimi K3	2.8T	1M	July 2026	$3 / $15
DeepSeek V4 Pro	1.6T	1M	April 2026	$0.46 / $1.07
Xiaomi MiMo-2	1.02T	256K	May 2026	$0.50 / $1.50
Kimi K2.6	1T	256K	Jan 2026	$0.95 / $4
Alibaba Qwen3-235B	235B	128K	April 2026	$0.40 / $0.80

K3 is in a completely different weight class. At 2.8 trillion total parameters, it nearly doubles the next-largest open model. Moonshot is marketing it as the “first open 3T-class model” — rounding up, but the gap is real. One nuance worth noting: like all MoE models, only a fraction of parameters are active per token. K3 activates 16 out of 896 experts, while DeepSeek V4 Pro activates 49B out of 1.6T (~3%). Raw parameter counts don’t tell the full compute story, but at this scale the total parameter capacity still matters for knowledge breadth and reasoning depth.

This isn’t just parameter-count flexing. The model’s architecture is genuinely novel, with two key innovations open-sourced on GitHub:

Kimi Delta Attention (KDA) — a hybrid linear attention mechanism that fundamentally changes how the model processes long sequences. Unlike standard attention where cost scales quadratically with context length, KDA enables efficient scaling to 1M tokens without the usual computational cliff. Moonshot contributed their vLLM implementation for production serving.
Attention Residuals (AttnRes) — a drop-in replacement for residual connections that selectively retrieves representations across depth instead of accumulating them uniformly. This delivers consistent scaling gains with fewer parameters.

Under the hood, K3 uses a Stable LatentMoE architecture — a Mixture of Experts design activating only 16 out of 896 experts per token. The routing uses Quantile Balancing (selecting experts based on router-score quantiles rather than simple top-k), which avoids the training instability issues that plagued earlier MoE models. Combined with Gated MLA (multi-head latent attention with improved selectivity), SiTU activations (Sigmoid Tanh Unit for better activation control), and Per-Head Muon (optimizing attention heads independently), Moonshot claims ~2.5× better scaling efficiency compared to K2.

Kimi K3 Benchmarks: vs GPT-5.6 Sol, Claude Opus 4.8 & DeepSeek V4

K3’s self-reported benchmarks position it firmly in the frontier tier, though it still trails the absolute top proprietary models:

Knowledge Work & Agentic Tasks

Benchmark	Kimi K3	Claude Fable 5 Max	GPT-5.6 Sol Max	Claude Opus 4.8
GDPval-AA v2 (44 real-world occupations)	1,668	1,815	1,748	1,600
AA-Briefcase (long-horizon agent)	1,548 (#2)	1,587	1,495	—
BrowseComp (information seeking)	91.2 (#1)	—	—	—

GDPval-AA v2 is particularly interesting — it’s a benchmark of 44 real occupational tasks, from software engineering to legal analysis. K3 places third behind only the two undisputed frontier leaders, and decisively ahead of Claude Opus 4.8. One notable gap: on DeepSWE (a major coding agent benchmark), K3 scored 67.5 vs GPT-5.6 Sol’s 73.0 — a 5.5-point delta that Moonshot’s self-selected benchmarks don’t highlight.

Coding Benchmarks

Benchmark	Kimi K3 Rank	Notes
SWE Marathon	#1	Agentic coding with real-world repos
Program Bench	#1	Competitive programming
Frontend Code Arena	#1	Human preference (Arena.ai) — beat Claude Fable 5
Automation Bench	#1	End-to-end task automation
SpreadsheetBench 2	#1	Complex spreadsheet reasoning

The Frontend Code Arena result is particularly striking — K3 topped the human preference leaderboard, surpassing even Claude Fable 5. This aligns with reports that K3 excels at “vision in the loop” workflows: generating code from screenshots and iterating on visual feedback.

Kimi K3 Pricing: Premium for a Chinese Lab

K3’s API pricing breaks from the Chinese-lab tradition of aggressive undercutting:

Model	Input ($/1M tokens)	Output ($/1M tokens)	Cached Input
Kimi K3	$3.00	$15.00	$0.30
DeepSeek V4 Pro	$0.46	$1.07	$0.11
Claude Sonnet 4	$3.00	$15.00	$0.30
Claude Opus 4.8	$10.00	$50.00	$1.50
GPT-5.6 Sol	$5.00	$40.00	$1.00

At $3/$15, K3 matches Anthropic’s Claude Sonnet pricing — making it the most expensive model ever released by a Chinese AI lab. DeepSeek V4 Pro, for comparison, is roughly 6-14× cheaper.

However, Artificial Analysis reports that K3 is surprisingly token-efficient: it produces 21% fewer output tokens than K2.6 for equivalent tasks. And their cost-per-task metric puts K3 at $0.94, close to GPT-5.6 Sol at $1.04 and half of Opus 4.8 at $1.80. So while the per-token price is high, the effective cost per completed task is competitive.

Moonshot is also offering a 30% voucher rebate on API credits of $1,000+ until August 12 — clearly targeting enterprise adoption.

The KDA Problem: Prefix Caching & Real-World Serving Costs

There’s a catch to the KDA architecture. As Simon Willison noted:

“KDA poses challenges for conventional prefix caching, which is how most production LLM deployments keep costs manageable.”

Standard Transformer attention is cache-friendly — the key-value pairs from earlier tokens don’t change when new tokens arrive, so they can be computed once and reused across requests. KDA’s hybrid linear attention breaks this property, meaning the standard KV-cache optimization doesn’t work out of the box. In practice, this means each request has to recompute attention from scratch, driving up serving costs. Moonshot has contributed a vLLM implementation to help, but expect real-world per-request costs to be higher than the raw token pricing suggests until the serving ecosystem adapts.

What Kimi K3 Can Do: Autonomous Agent Demos

Moonshot’s technical blog showcases genuinely impressive agentic capabilities that go beyond benchmark tables:

48-Hour Autonomous Chip Design

K3 was given open-source EDA tools, the Nangate 45nm library, and a goal: design a chip capable of running a nano-scale version of itself. After 48 hours of continuous autonomous operation, it produced:

A 4 mm² chip closing timing at 100 MHz
1.46M standard cells, 0.277MB SRAM
INT4 MAC array with fused dequantization
8,700+ tokens/s decode throughput in simulation

This wasn’t a scripted demo — K3 read documentation, made architectural decisions, ran verification loops, and iterated on failures. In the chip design world, this is roughly equivalent to a junior engineer’s 2–3 month internship output, compressed into two days.

Kernel Optimization: The Model Optimizing Itself

In the late stages of K3’s development, the Moonshot team used an early version of K3 to handle the majority of their GPU kernel optimization work. The model autonomously profiled, rewrote, and benchmarked kernels for AttnRes, KDA, and MLA across NVIDIA H200s and alternative GPGPUs, producing code competitive with Claude Fable 5 and substantially outperforming GPT-5.6 Sol.

Computational Astrophysics in 2 Hours

K3 was tasked with reproducing the I-Love-Q universal relation — a problem that typically takes an experienced researcher 1–2 weeks. It reviewed and cross-validated 20+ papers, identified formula inconsistencies across sources, implemented a full numerical pipeline in 3,000+ lines of Python, evaluated 300+ equations of state, and produced an interactive HTML dashboard. Total time: ~2 hours.

The Backstory: From Hype to Pivot

Moonshot AI’s trajectory is a case study in the brutal pace of the AI industry:

2023: Founded by Yang Zhilin (Tsinghua PhD, ex-Google Brain, ex-Meta AI)
2024: Raised ~$1.5B at a $4.3B valuation; Kimi platform popular for long-text analysis in China
Jan 2025: DeepSeek R1 drops. Kimi’s monthly active users in China drop from 3rd to 7th as users flock to DeepSeek’s free tier
July 2025: Pivot to open-source with Kimi K2
Jan 2026: Kimi K2.6 (1T params)
July 2026: Kimi K3 — culmination of the open-source strategy

The pivot was existential. DeepSeek’s aggressive open-source + ultra-low pricing strategy decimated Kimi’s user base. K3 is Moonshot’s answer — not by competing on price, but by being the biggest, most capable open model on the market.

The Geopolitical Context

K3’s timing — one week before the World AI Conference in Shanghai — is no accident. China’s AI labs are locked in an escalating open-source arms race:

Kimi K3:     ████████████████████████████████ 2.8T
DeepSeek V4: ██████████████████ 1.6T
Xiaomi MiMo: ████████████ 1.02T
Qwen3:       ████ 235B
K2.6:        ████████████ 1T

Each release leapfrogs the last. The strategic calculus is clear: open-source models erode the moat of proprietary frontier labs. When a free, downloadable model performs at 90–95% of GPT-5.6 Sol, the $200/month ChatGPT Pro subscription becomes a harder sell.

Kimi K3 Recommendations: Should You Use It?

✅ Best For

Agentic coding with visual feedback — K3 is currently #1 on Frontend Code Arena
Long-horizon autonomous tasks — 48-hour chip design, multi-day research pipelines
Open-source deployments — full weights drop July 27, deploy on your own infrastructure
Knowledge work with 1M context — ingest entire codebases, research papers, or documentation

⚠️ Consider Alternatives For

Budget-constrained projects — DeepSeek V4 Pro is 6× cheaper with comparable quality
Latency-sensitive applications — KDA makes prefix caching harder, expect higher TTFT
Simple tasks — $15/M output tokens is overkill for classification or extraction

The Pelican Test

Simon Willison ran his famous pelican-riding-a-bicycle SVG test on K3. The result? A “cartoon illustration of a white pelican wearing a red scarf, riding a red bicycle” that cost 25 cents for a single generation (13,241 reasoning tokens + 3,417 output tokens). The reasoning overhead is significant — K3 currently only supports maximum reasoning effort, with no way to dial it down for simple tasks.

The Bottom Line

Kimi K3 is a shot across the bow. At 2.8 trillion parameters, it’s the largest open-weight model ever released — and its benchmark performance justifies the scale. It doesn’t beat Claude Fable 5 or GPT-5.6 Sol, but it gets close enough that the difference becomes a business decision rather than a capability gap.

For developers, the July 27 weight release is the real prize. A 2.8T-parameter model you can run on your own infrastructure, fine-tune for your domain, and deploy without API rate limits — that’s the kind of leverage that changes how teams build.

The open-source frontier model gap is now measured in weeks, not months. And with K3, it just got a lot smaller.

Full benchmark data: kimi.com/blog/kimi-k3
API access: platform.kimi.ai
GitHub: github.com/moonshotai