On July 16, 2026, the Chinese company Moonshot AI (月之暗面) released Kimi K3 — its most powerful model to date. This is a sparse Mixture-of-Experts (MoE) system with 2.8 trillion parameters, a context window of 1 million tokens, and a hybrid attention architecture designed for efficient operation on long sequences. The model is positioned as a serious competitor to Claude Opus 4.8 and GPT-5.5, and on specialized tasks (frontend coding, agent systems, repository work) it demonstrates results that are on par with or exceed these closed systems.
Full weights are promised to be released on July 27, 2026. If the release happens as announced, Kimi K3 will become one of the largest open-weight models ever released, surpassing its predecessors in scale by more than 2.5 times.
Company and Strategic Context
Moonshot AI is a Beijing-based lab founded in 2023 by Yang Zhilin and his co-founders from Tsinghua University, known for its open strategy and the Kimi series of models. The company has received significant investment: in May 2026, it closed a $2 billion round at a $20 billion valuation, backed by Meituan's Long-Z Investments and other key investors including Alibaba and Tencent. Quarterly ARR exceeded $200 million in April 2026. The Kimi series has steadily grown, from K2 (1T parameters) through K2.5–K2.7 Code to the current K3, each time expanding context and adding multimodality. K3 represents a qualitative leap in both scale and architectural innovation, specifically targeting long-tail agent tasks and coding.
Architectural Innovations: The Core of Competitive Advantage
1. Kimi Delta Attention (KDA) — a hybrid linear attention mechanism, evolving ideas from Gated DeltaNet and previously tested in the open model Kimi Linear (2025). Instead of the classic quadratic attention, which requires O(n²) memory and computation, most layers use a linear complexity scheme with channel-wise diagonal gating. Practical benefits:
- Reduced KV cache consumption by up to 75% in similar architectures, which is critical for handling million-token contexts;
- Decoding speedup in million-token contexts by up to 6.3×, making long sessions practically deployable;
- Preservation of retrieval quality on long sequences without significant degradation.
KDA has received open-source CUDA kernels (FlashKDA) and is integrated into vLLM for production deployments.
2. Attention Residuals — a mechanism that selectively “pulls” representations from earlier layers instead of uniform accumulation through standard residual connections. Moonshot reports approximately a 25% increase in training efficiency with minimal additional cost (~2%), which, at the scale of a 2.8T parameter model, means significant savings in computational resources for training or an equivalent increase in quality.
3. Stable LatentMoE — an extremely sparse routing system where out of 896 experts, only 16 are activated per token (1.8% of the pool), using quantile balancing and per-head optimization for training stability. This level of sparsity allows inference computational cost to be kept at an acceptable level despite the enormous number of parameters: due to the MoE architecture and 16 active experts, the actual number of activated parameters per token is significantly lower than 2.8T.
4. Quantization and Efficiency — the model is initially quantized using MXFP4 for weights and MXFP8 for activations, chosen for compatibility with diverse hardware. Moonshot claims approximately 2.5× better scaling efficiency compared to the K2 generation, meaning the same amount of computation provides significantly more capabilities.
Technical Specifications
| Parameter | Value |
Total Parameters | 2.8 trillion |
Active Experts | 16 out of 896 (1.8% per token) |
Context Window | 1,048,576 tokens (1M) |
Modalities | Text + Image/Video → Text (native vision) |
Core Architecture | Sparse MoE + Kimi Delta Attention + Attention Residuals |
Quantization | MXFP4 (weights) / MXFP8 (activations) |
Reasoning | Always on (“thinking mode”), max-effort only at start, low/high modes planned |
Independent and Official Benchmarks: The Real Performance Picture
Leadership in Specialized Tasks: On the Frontend/WebDev Arena (developer human preferences), K3 took 1st place with 1679 Elo, surpassing Claude Fable 5 (1631) and GPT-5.6 Sol (1618) in blind testing. This is a 17-position jump from K2.6's rank and the #1 result in 6 out of 7 frontend domains (excluding games).
Overall Performance: On the Artificial Analysis Intelligence Index v4.1 (independent evaluation aggregating 9 benchmarks: GDPval-AA, reasoning, coding, agent tasks), K3 scored 57.1 — 4th place globally, behind Fable 5 (~60) and GPT-5.6 Sol (~59), but ahead of Claude Opus 4.8 (~56).
Coding and Agent Systems:
- Terminal-Bench 2.1: 88.3 (nearly on par with GPT-5.6 Sol 88.8, noticeably ahead of Claude Fable 5 and Opus 4.8 — both 84.6).
- GDPval-AA v2: 1668 Elo, third place after Fable 5 Max (1815) and GPT-5.6 Sol Max (1747.8), ahead of Opus 4.8 (1600) — a standard test for knowledge, reasoning, and analytical tasks.
- GPQA-Diamond: 93.5% — the best open result on this high-difficulty benchmark (PhD-level knowledge).
- BrowseComp (Web Agents): 91.2% (best result among all compared models), 90.4% on the full 1M context without compression.
- AA-Briefcase (Long-term Knowledge/Agents): 1548 Elo, second place after Fable 5, ahead of GPT-5.6 Sol.
Some independent testers note increased “talkativeness” in the output, which could increase token consumption, but for core coding and agent tasks, output token usage is 21% lower than K2.6, thanks to reasoning token optimizations.
Pricing and Availability
API Pricing: $3.00 per million input tokens (cache miss), $0.30 for cached input, $15.00 per million output — the most expensive among all Chinese labs, but comparable to Claude Sonnet and approximately half the cost of GPT-5.6 Sol ($30/M output). For comparison, K2.6 cost $0.60 input / $2.50 output, so this is a 5× jump for input and 6× for output, reflecting a significant leap in scale and quality.
Availability: Kimi Code, Kimi app (including iOS), Moonshot OpenAI-compatible API, OpenRouter. Full 1M context requires an Allegretto subscription or higher; Moderato is limited to 256K.
Promotional Campaign: The launch was accompanied by a promotion with bonus credits of +10–30% on top-ups from July 15 to August 11, 2026.
Open-Weight Release: Expectation and Significance
At the time of launch (July 16), the weights were not yet published. Moonshot has officially committed to releasing the full weights on July 27, 2026 under a license close to Modified MIT (like the K2 series). Once the weights are opened, K3 will become the most powerful freely available model in terms of combined scale (2.8T), context (1M), and coding/agent capabilities, meaning it will be accessible for local development, fine-tuning, and research on appropriate hardware (minimum 64 accelerators for the full model).
Strategic Importance and Competitive Context
Kimi K3 arrives at a critical turning point in the Chinese AI industry. After DeepSeek's surge in January 2025 (R1 at low-cost computing), Moonshot lost market position, dropping from 3rd place in active Kimi users to 7th by June 2025. K3 is a deliberate bet to regain ground through scale and architectural innovation. Investors call this a potential “Kimi moment” (akin to the shock from DeepSeek a year ago), and there are grounds for it: the model closes the gap with Western frontier models not through cheapness (it's expensive), but through quality and scale.
The combination of 2.8T parameters, 1M context, native vision, always-on reasoning, and strong coding makes K3 particularly interesting for:
- Agent systems and long-horizon workflows (reproducing user interfaces, multi-step development, autonomous engineering tasks);
- Development and refactoring of large codebases (in one notable example, K3 autonomously designed a microchip in 48 hours using open-source EDA);
- Research and analytical tasks with vast context (legal documents, financial reports, code reviews of large repositories);
- Companies seeking self-hosting and independence from closed APIs rather than relying on Anthropic or OpenAI.
However, important caveats remain: the weights are not yet released, independent reproductions and detailed technical reports are still in development, and some benchmarks use proprietary harnesses (KimiCode instead of standard ones). The full picture will only become clear after July 27, when the weights become available and independent evaluations of open versions of the models emerge.
Conclusion: Kimi K3 is one of the most significant AI releases of the first half of 2026. It is not just “another big Chinese model,” but a serious contender for a place at the forefront of frontier models, especially in specialized tasks of agent coding, long contexts, and multimodal reasoning. Architectural innovations (KDA, Attention Residuals, Stable LatentMoE) demonstrate that Chinese labs are on par with the best Western ones in engineering. If the open-weight release proceeds as planned on July 27, the impact on the open-source ecosystem and competitive dynamics will be very significant: for the first time, developers will have access to a model of such scale and quality independent of closed APIs.
