Kimi K3: Release Date, Benchmarks, Open Weights, Market Impact

Kimi K3 is Moonshot AI’s 2.8-trillion-parameter open-weight model — the largest ever released as of July 2026. Since its launch on July 17, 2026, it has topped the LMArena Code Arena leaderboard, triggered a semiconductor selloff across Asian markets, overwhelmed Moonshot’s own GPU infrastructure, and drawn Microsoft’s attention as a potential Copilot backend. This page tracks the full story as it develops.

Live Updates: Kimi K3 Timeline

🟢 July 21, 2026 — Microsoft Is Quietly Testing Kimi K3 for Copilot

Reports emerged that Microsoft has begun evaluating Kimi K3 as a potential backend for Copilot, its flagship AI assistant. The driver is straightforward: at $3 per million input tokens with cache hit rates above 90%, Kimi K3 could save Microsoft hundreds of millions annually on inference costs currently going to OpenAI.

The evaluation is not a casual experiment. Microsoft engineers are running structured tests across Copilot’s core workflows. If results hold, this would represent the most significant challenge to OpenAI’s commercial relationship with Microsoft since the partnership began.

Azure already hosts third-party models through Foundry, including earlier Kimi versions, so adding K3 to Azure is relatively straightforward. Running the open weights on Azure-owned compute also keeps enterprise data inside Microsoft’s environment. The widely cited $600 million savings figure is an estimate, not a Microsoft-confirmed number.

🔴 July 20, 2026 — GPU Capacity Overwhelmed, New Subscriptions Paused

Forty-eight hours after launch, Kimi K3 demand had pushed Moonshot’s infrastructure past its limits. The company announced via X that new subscriptions were temporarily paused to protect existing users from degraded performance.

Moonshot responded with a structural fix: splitting its single membership into two tiers. The standard Kimi Membership covers web, app, and work use cases. The new Kimi Code Membership dedicates compute specifically to agentic coding workflows — the category driving the heaviest load. A subscription pause this fast is rare in AI. It signals genuine product-market fit, though it also created friction for developers who needed immediate access.

Third-party gateways, including ZenMux and Zyloo, stepped in to offer free access windows while Moonshot expanded capacity.

📈 July 17–18, 2026 — Markets React, Searches Surge 500%

The launch triggered immediate market movement. Google Trends data showed “Kimi K3” searches rising 450% and “Kimi AI” up 350% within hours of the announcement. Analysts drew instant comparisons to DeepSeek’s January 2025 debut, which cost Nvidia $590 billion in market cap in a single session.

This time the damage was more distributed. Taiwan’s benchmark index fell over 6%, Japan’s market dropped 4%, and the Nasdaq slid 1.5%. Semiconductor ETFs fell below key support bands as investors reassessed the economics of proprietary AI infrastructure.

The core concern is structural: if a Chinese open-weight model delivers frontier performance at $3 per million tokens, the premium infrastructure spending assumptions that underpin Western AI valuations start to look fragile. Chinese models already account for 45% of OpenRouter traffic. Kimi K3 raised the question of how far that number can still go.

🥇 July 17, 2026 — LMArena Code Arena: Kimi K3 Takes #1

Within hours of launch, Kimi K3 debuted at the top of LMArena’s Code Arena — the most credible human-preference coding leaderboard available. It displaced GPT-5.6, Claude Sonnet, and Gemini 2.5 Pro. The margin was most pronounced in the frontend development category, which was previously considered a stronghold for American models.

LMArena’s methodology matters here. Unlike static benchmarks, Code Arena uses blind preference voting from real developers given identical prompts. There are no leaked test sets and no prompt-tuning shortcuts. Reaching number one reflects genuine utility as judged by professional coders — not a marketing benchmark.

Frontend output demands more than syntactic correctness: semantic HTML, responsive CSS, bug-free interactivity, and visual taste. Kimi K3 scoring highest on human preference in this category suggests it has internalized frontend judgment, not just syntax rules.

🚀 July 17, 2026 — Official Launch: 2.8T Parameters, Custom Chip, Open Weights

Moonshot AI officially launched Kimi K3 at 00:00 CST on July 17. The confirmed parameter count — 2.8 trillion — exceeded pre-launch leaks that had estimated 2.5 trillion, making this the largest open-weight model ever released. DeepSeek V4 Pro previously held that record at 1.6 trillion parameters. Kimi K3 nearly doubled it.

What Kimi K3 Actually Is

Kimi K3 is a Mixture-of-Experts model that activates 16 of 896 experts per token, keeping inference costs manageable despite the headline parameter count. Moonshot claims 2.5x scaling efficiency over Kimi K2 — meaning it converts compute into capability at a significantly higher rate than its predecessor.

Architecture Innovations

Two custom components define how the model processes information. Kimi Delta Attention improves long-sequence handling and required a new vLLM implementation to support prefix caching effectively. Attention Residuals selectively retrieve representations from earlier layers, improving information flow across the model’s depth.

Training uses quantization-aware methods with MXFP4 weights for broad hardware compatibility. Stable LatentMoE keeps routing stable through Quantile Balancing and Per-Head Muon optimization. Moonshot recommends supernode deployment configurations with high-bandwidth communication for best results.

Benchmarks: Where It Leads and Where It Trails

Kimi K3 leads on agentic and coding evaluations. It scores 81.2 on FrontierSWE, ahead of GPT-5.6 Sol’s 71.3. It also outperforms competitors on SWE Marathon, BrowseComp, and DeepSearchQA. Hard reasoning benchmarks like HLE-Full remain a gap — Claude Fable 5 and GPT-5.6 Sol hold advantages there. Overall, Kimi K3 sits comparable to Claude Opus 4.8 in aggregate scoring.

One caveat applies to Moonshot’s comparison against Claude Fable 5 on coding tasks. Moonshot notes potential fallback behavior during evaluation that may affect that specific comparison.

What the Demos Actually Showed

Moonshot published case studies that went beyond benchmark tables. In one test, the model optimized GPU kernel performance by 60% over 15 hours without changing numerical correctness. It also built MiniTriton — a compact GPU compiler — from scratch, performing on par with Triton on roofline benchmarks.

The hardware demo was the most striking: Kimi K3 designed a functional chip for a nano model in 48 hours using open-source EDA tools, producing a design that closed timing at 100 MHz. Community testing added to the picture. Developer Max Weinbach used a single prompt to build a working macOS 27 simulator, and a separate demo produced a functional browser OS from one instruction. Both circulated widely and contributed to the search surge.

Native Vision and Multimodal Capabilities

Vision capabilities are built directly into the base model rather than added as an external layer. This enables complex scientific visualization tasks, interactive financial reports generated through hundreds of self-improvement rounds, and procedural 3D game environments built in single sessions. The “vision in the loop” design allows iterative feedback between code generation and visual output — genuine multimodal agentic workflows rather than simple image input.

Pricing and Access

The model is available via the Kimi API and desktop app. Cache-hit input costs $0.30 per million tokens; output is $15.00. Moonshot reports cache hit rates above 90% in coding workloads, making effective input costs significantly lower than list rates suggest. Full open weights release on July 27 under a modified MIT license — attribution required only for products exceeding 100 million monthly active users.

Third-party access is available through Forge AI, which offers free credits to sample Kimi K3 alongside other frontier models with no credit card required.

Open weights: download, size, and hardware

Moonshot released the full Kimi K3 weights on July 27, 2026, at huggingface.co/moonshotai/Kimi-K3. Download only from that verified organization. The native release is MXFP4 safetensors at roughly 594 GB. Community GGUF and BF16 re-quantizations are expected to follow, with Q4 variants likely landing in the 300 to 400 GB range.

The weights include native vision and a one-million-token context window. Moonshot also contributed a Kimi Delta Attention prefill cache implementation to upstream vLLM so the model loads reliably across inference stacks.

Hardware needed to self-host

Kimi K3 does not run on a consumer GPU. The minimum experimental configuration is about 8x H100 80GB GPUs. Moonshot recommends supernode configurations with 64 or more accelerators for production throughput. An RTX 4090, Mac Studio, or similar machine cannot load the model even with quantization.

As of the release, AWS Bedrock, Microsoft Azure Foundry, and Google Vertex AI had not added Kimi K3 as a managed offering, according to Bloomberg Intelligence analysts. Third-party open-model hosts were expected to benefit first.

The limitations Moonshot published

Moonshot included three unusual admissions in its technical blog. First, K3 was trained in preserved thinking history mode, so quality can become unstable if an agent harness drops earlier thinking content or switches models mid-session. Second, the model can be excessively proactive on ambiguous tasks and may make decisions on the user’s behalf unless the system prompt sets boundaries. Third, Moonshot concedes a noticeable user-experience gap compared with Claude Fable 5 and GPT-5.6 Sol.

Washington scrutiny and the China data question

White House OSTP Director Michael Kratsios accused Moonshot of training K3 on banned Nvidia chips and distilling US models, including Anthropic’s Fable. The allegations are unconfirmed and Moonshot did not comment. They still matter to US enterprises evaluating K3 for production.

Moonshot is a Beijing-based company subject to China’s National Intelligence Law. Hosted API traffic passes through servers under that framework. Self-hosting the open weights removes that exposure, which is the main reason a regulated company would absorb the hardware cost.

Open weight is not open source

Moonshot markets K3 as open source, but the accurate term is open weight. Buyers get the trained parameters, which can be downloaded, fine-tuned, and deployed commercially under a modified MIT license. Training data and the full training pipeline are not included.

Background: How Moonshot Got Here

Yang Zhilin founded Moonshot in 2023 after working at Google Brain. By early 2025, Kimi ranked third among Chinese AI platforms. DeepSeek’s R1 launch in January 2025 pushed Moonshot to seventh. The company responded by pivoting deliberately toward open-source development. Kimi K2.6 topped open-weight leaderboards by April 2026, and Moonshot raised $2 billion in May 2026 at a $20 billion valuation — bringing total funding to $3.77 billion. Annual recurring revenue reportedly crossed $200 million in April 2026. Kimi K3 is the payoff on that rebuild.

What to Watch Next

Three things will determine Kimi K3’s actual competitive impact. First, community adoption after the July 27 weight release. Download volume and fine-tune activity will show whether self-hosting turns into real usage. Second, the Microsoft evaluation outcome — if Copilot switches any workload to Kimi K3, it reframes the entire US-China AI competitive narrative. Third, the next benchmark cycle — whether the LMArena lead holds as other labs update their models will show whether this performance is durable or a launch-window spike.

Related Reading

• Kimi K3 Demand Crushed Moonshot’s GPU Capacity in 48 Hours
• 5 Things Kimi K3 Can Do That Most Users Haven’t Tried Yet
• July 2026 Was the Wildest Month in AI History

Frequently Asked Questions

When was Kimi K3 released?

Kimi K3 launched on July 17, 2026. Moonshot released the full model weights on July 27, 2026.

What is Kimi K3?

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts language model developed by Moonshot AI. It launched on July 17, 2026, as the largest open-weight AI model ever released, with native vision, a one-million-token context window, and agentic coding capabilities.

When did Kimi K3 open weights release?

Moonshot released the full weights on July 27, 2026, under a modified MIT license. Products with fewer than 100 million monthly active users can use them without attribution.

How does Kimi K3 compare to GPT-5.6 and Claude?

Kimi K3 leads GPT-5.6 Sol on FrontierSWE (81.2 vs 71.3) and topped LMArena’s Code Arena for frontend development. It trails on hard reasoning benchmarks like HLE-Full, where Claude Fable 5 and GPT-5.6 Sol hold an advantage. Overall capability is comparable to Claude Opus 4.8.

Why did Kimi K3 cause a market selloff?

The launch raised investor concerns that frontier AI performance no longer requires premium Western infrastructure. If open-weight Chinese models close the capability gap at lower cost, the economics underpinning Nvidia and US AI infrastructure spending weaken. Taiwan’s market fell over 6% and Japan’s dropped 4% in the immediate aftermath.

Is Kimi K3 free to use?

The API is paid, at $0.30 per million cache-hit input tokens and $15.00 per million output tokens. Third-party gateways including Forge AI have offered free access credits. The open weights released on July 27 make self-hosting possible, and the license keeps attribution optional for smaller products.

How big is the Kimi K3 download?

About 594 GB for the native MXFP4 safetensors release. Community Q4 quantizations are expected to land around 300 to 400 GB.

What hardware do I need to run Kimi K3 locally?

At least 8x H100 80GB GPUs for experimental inference, and 64 or more accelerators for production throughput. Consumer GPUs cannot load the full model.

Is Kimi K3 open source?

It is open weight, not open source. The parameters are downloadable and commercially usable under a modified MIT license, but the training data and training pipeline are not released.