Open-Source Giant Model Weights Released: 2.8T-Parameter MoE Architecture and Agent Capabilities Set a New Benchmark

Architecture: Balancing Sparse Expert Networks with Long-Context Attention

The model employs a 93-layer Transformer architecture, but its core design strength lies in a modular reorganization of attention mechanisms and expert networks. Specifically, 69 layers use KDA (Key-Value Decomposed Attention), while the remaining 24 layers are upgraded to Gated DSA (Gated DeepSeek Sparse Attention). This hybrid design is not an arbitrary stack; it represents a precise trade-off between memory footprint and computational efficiency for long-sequence modeling.

Compared with the previous generation, the most significant structural change is the full replacement of the all-attention Gated MLA (Multi-head Latent Attention) layers with Gated DSA. The rationale: when processing contexts exceeding one million tokens, full attention consumes memory quadratically, while sparse attention uses gating logic to dynamically select critical tokens, reducing computational complexity from quadratic to near-linear. This makes ultra-long-context engineering feasible. During inference, this design significantly reduces KV Cache usage, leading to more stable response times in multi-turn dialogue and long-document parsing.

On the expert network side, the model configures 896 fine-grained experts; each token activates only 16 of them, plus 2 always-on shared experts. This “fine-grained + shared expert” strategy substantially increases parameter capacity without a proportional rise in compute cost. The roughly 27x gap between 2.8T total parameters and 104B activated parameters directly reflects this sparse activation property—the model is large, but each forward pass is computationally equivalent to a 10-billion-parameter dense model.

Quantized Storage and Training: Performance and Precision Trade-offs with MXFP4/8

To address the memory bottleneck in large-scale deployment, the official release uses a quantization-aware training scheme. Weights are stored on disk and in GPU memory in MXFP4 (Microscaling Floating Point 4-bit) format, while activation values are computed in MXFP8 (8-bit) precision during the forward pass. This asymmetric quantization strategy is pragmatic: weights are static, and 4-bit precision preserves parameter distribution characteristics; activations are dynamic, and 8-bit precision avoids precision loss from fluctuation in activation magnitudes.

Through the Hugging Face repository interface, the full repository occupies approximately 1.56TB, with the parameter count estimated at about 2.78 trillion based on safetensors header metadata. MXFP4 storage reduces weight volume by roughly 75% compared with FP16 or BF16, which is the key enabler for fitting a 2.8T model onto a single 8-GPU high-performance server. Notably, this quantization is not a post-training compression step; it is quantization-aware training (QAT) deeply integrated into the training pipeline. The model adapts to low-bit noise during gradient backpropagation, keeping inference precision loss minimal. For developers, this means the downloaded weights can be used directly with good inference quality, without requiring additional calibration or quantization.

Deployment Requirements and Hardware Estimates: Practical Constraints for Enterprise Private Hosting

Despite aggressive quantization, deployment still demands top-tier hardware. Based on estimates from open-source community enthusiasts, a minimal viable configuration is a server with 8 NVIDIA B300 (Blackwell Ultra) GPUs, each with about 288GB of VRAM, totaling roughly 2.3TB. At current market prices, hardware cost alone is estimated between $550,000 and $650,000. This is a third-party estimate; the official team has not published baseline hardware recommendations. Real-world deployment also requires attention to CPU memory bandwidth, NVLink interconnect bandwidth, and high-speed storage.

This hardware threshold defines the target audience—not individual developers, but AI labs and enterprises with clear business needs and infrastructure budgets. For those without such budgets, a more practical route is calling the cloud API. The official API model identifier is kimi-k3.1-pro, compatible with both OpenAI and Anthropic API protocols, reducing migration friction. An accelerated version, kimi-k3.1-pro-highspeed, is also available; the provider claims 2–3x throughput of the standard version, suited for latency-sensitive real-time interactions. Developers can control reasoning depth via the reasoning_effort parameter (options: low, high, max; default is max).

Agent Capability Evidence: Quantitative Evaluation from Real-World Sessions

The central metric for current large models has shifted from text generation to complex task planning and tool invocation. According to the authoritative Agent Arena evaluation platform, this model performs strongly in long-horizon agent sessions with real users. The platform calculates a “net lift” score based on multiple signals: explicit user confirmation of task success, praise or complaints about model output, model correction responses after user error reports, and whether Bash command execution succeeds.

On the day of the weight release, the model (with Max reasoning configuration) achieved a net lift of +5.1%, ranking first among open-weight models and 13th overall. This is a decisive improvement over the previous generation, which posted only +0.91% net lift under Max configuration and ranked sixth among open-weight models. In sub-scores, the model topped the open-weight leaderboard across Work, Code, and Chat categories. Notably, on Bash reliability—measuring terminal operation accuracy—the model reached +7.9%, a qualitative leap from the prior generation’s -4.5%, demonstrating high stability in tool invocation and command execution on real server environments.

Inference Stack Ecosystem and Toolchain Support: Three Mainstream Options

Open-source inference ecosystems have already adapted to this model. The mainstream engines—vLLM (high throughput with PagedAttention memory management), SGLang (structured generation and multimodal support), and TokenSpeed (low-latency optimization)—have all published deployment recipes for the model. Developers do not need to write custom CUDA kernels or fusion logic; they only need to specify the model path and quantization format in the inference framework’s configuration file to use optimized high-concurrency execution. This ecosystem synchronization significantly shortens the path from weight download to service launch.

Multimodal and Vision Understanding: The MoonViT-V2 Vision Tower

To handle mixed text-image input, the model includes a dedicated vision encoder module: MoonViT-V2 (a Vision Transformer) with 401M parameters. This module maps image pixels into semantic vectors aligned with the text space, enabling the model to process scanned documents, charts, screenshots, and other complex visual inputs. In scenarios such as visual question answering, UI automation testing, and OCR for mathematical formulas, the coordination between this vision tower and the language backbone directly affects output quality.

Conclusion and Selection Guidance

This release marks a new equilibrium between parameter scale and practical performance for open-weight models. The million-token context window combined with sparse attention makes it an excellent foundation for long-document analysis and codebase-level code generation. The 2.8T sparse parameter capacity provides substantial knowledge reserves for complex reasoning. For institutions with sufficient budgets, private deployment ensures data privacy and customization freedom through fine-tuning. For rapid experimentation or elastic demand, API calls offer clear cost and iteration advantages. When making a business decision, the key factor is your actual requirement level for reasoning_effort: higher reasoning strength improves success on complex tasks but incurs greater latency and cost. A sensible load-balancing strategy will be essential for production deployment.