Open-Weight 2.8T-Parameter Model Sets a New Benchmark for Agent-Capable LLMs

From Parameter Scale to Emergent Capability: A New Benchmark for Open Large Models

While the industry continues to debate the operational cost of hundred-billion-parameter models, a frontier AI research organization followed through on its commitment in early November and released the complete weight files of its flagship large model to the community. This is not merely another open-source release. It marks the entry of ultra-large-scale models, at the 2.8-trillion-parameter level, into a phase where they can be downloaded, studied, and deployed privately. The model not only has a striking parameter count but also sets a new score record among open-weight models for agent task execution. This article takes a close look at the technical foundation, deployment challenges, and real-world performance of this model, providing a practical reference for teams evaluating ultra-large-scale model technology.

1. Hardware Specifications and Architectural Innovations: More Than Just Scale

With a total of 2.8 trillion parameters, the internal structure deserves careful examination. The model uses the standard Mixture of Experts (MoE) architecture, a mainstream approach for significantly increasing model capacity without proportionally raising compute cost.

• Parameter distribution and activation mechanism: The total parameter count is approximately 2.78 trillion, but not all parameters are invoked during every inference step. Each input token activates only about 104 billion parameters. This means the actual compute required for inference is far lower than a dense model of equivalent scale. This sparse activation property is central to maintaining a high level of capability while keeping serving costs under control.
• Layer stacking and attention innovations: The model consists of 93 layers, including 69 Key-Value Decomposed Attention (KDA) layers and 24 Gated DeepSeek Sparse Attention (Gated DSA) layers. Compared with the previous generation, the most significant change is in long-context modeling: the fully attentive Gated MLA layers have been replaced with Gated DSA layers. The introduction of sparse attention addresses the quadratic compute growth associated with long-text processing, making resource consumption more manageable when handling context windows of up to one million tokens.
• Expert routing mechanism: The model deploys 896 experts in its expert layers. During each forward pass, the system dynamically selects the 16 most relevant experts based on the features of the current token, while 2 shared experts always participate in computation. This fine-grained routing strategy helps the model switch flexibly across knowledge domains and improves the precision of knowledge retrieval.
• Multimodal vision capability: Unlike text-only models, this system includes a built-in vision encoder, MoonViT-V2, with 401 million parameters. This gives the model native image-text understanding and the ability to process interleaved visual and textual inputs, expanding its applicability in document analysis, embodied intelligence, and related scenarios.

2. Storage and Quantization Strategy: Balancing Precision at 1.56TB

For a 2.8T-parameter model, efficiently storing and loading the weights is one of the first major engineering challenges. The release adopts a forward-looking quantization-aware training approach.

• MXFP4 storage and MXFP8 activations: Weights are compressed for storage using MXFP4, a 4-bit floating-point micro-scaling format, while activations during forward propagation are kept in MXFP8, an 8-bit floating-point micro-scaling format. This is a typical low-bit storage, high-precision compute strategy. Through quantization-aware training, the model simulates the information loss caused by low-precision storage during training, minimizing performance degradation when the released low-precision weights are used for inference.
• Actual size and parameter statistics: This aggressive quantization strategy achieves an impressive compression ratio. The model occupies approximately 1.56TB in its Hugging Face repository. Stored in conventional FP16 format, the size would be nearly four times larger. Inspecting the safetensors file headers confirms a parameter count of roughly 2.78 trillion, consistent with the officially announced 2.8T total, confirming that quantized storage does not reduce the structural scale of the model.

Deployment ecosystem compatibility: For teams planning to self-host the model, mainstream inference frameworks have moved quickly. Widely used high-performance inference engines such as vLLM, SGLang, and the newer TokenSpeed provided deployment recipes on the release day, with support for tensor parallelism and pipeline parallelism to enable multi-GPU distributed inference.

3. Hardware Requirements and Deployment Reference

Although the weights are openly available, the physical hardware barrier remains high. For organizations with substantial compute resources, local deployment is feasible.

• Memory and compute estimates: Based on the 1.56TB model size, loading the full model into memory requires at least 2TB of aggregate GPU memory. According to an estimate from a community technical enthusiast (@kimmonismus), one viable minimum configuration is a server with eight NVIDIA B300 GPUs, each expected to offer 288GB of memory, for roughly 2.3TB total. Given the current premium pricing of B300 accelerators, the estimated hardware cost for this configuration ranges from $550,000 to $650,000, covering server hardware only and excluding data center space and power.
• Inference throughput trade-offs: While loading the model is theoretically possible, real-world inference speed will be constrained by memory bandwidth and expert-parallel communication overhead. Supporting high-concurrency production workloads typically requires multi-node tensor parallelism, such as two 8-GPU servers, which further increases hardware cost.

4. Agent Capability Evaluation: Measurable Gains in Real Sessions

Parameter count is only a static metric. The real test lies in task execution capability. The model’s performance on the third-party Agent Arena benchmark provides compelling evidence.

• A benchmark with a distinctive methodology: Agent Arena is not a traditional static question-answering test set. It collects anonymous logs from real users interacting with models in multi-turn agent tasks such as writing code, executing shell commands, and operating a computer. By analyzing signals including explicit user confirmation of task success, praise and complaints, model responses after correction, and whether shell commands executed successfully, it calculates a net lift score. This mechanism more realistically reflects a model’s reliability and usefulness in solving open-ended problems.
• Dominant open-model performance: Based on more than 17,500 real long-duration sessions, the model in its Max inference configuration achieved a net lift rate of +5.1%, ranking first among open-weight models and 13th overall. In the Work, Code, and Chat subcategories, it ranked first among open models in all three. By comparison, the previous-generation model scored only +0.91%, ranking sixth, highlighting a substantial generational improvement in agent task capability.
• Engineering detail: reliability of Bash tool calls: In Bash tool invocation, the most common scenario in agent tasks, the model achieved a reliability score of +7.9%, also first among open models. The previous generation scored -4.5% on this metric, meaning it frequently failed when attempting complex shell commands due to hallucination or syntax errors, degrading the overall user experience. This improvement demonstrates a qualitative leap in instruction following and structured output.

5. API Access for Faster Development

For developers unwilling to bear the high hardware cost, cloud API access offers an efficient alternative for evaluating the model’s capabilities.

• API model naming and compatibility: The API model name is kimi-k3.1-pro, with full compatibility with both the OpenAI and Anthropic interface protocols. Existing agent frameworks or applications built on these protocols can switch by changing only the base URL and API key, without modifying core code logic.
• Performance-enhanced variant and parameter tuning: An accelerated endpoint, kimi-k3.1-pro-highspeed, is also available, with response speeds reported to be two to three times faster than the standard version. Developers can use the reasoning_effort parameter, with options low, high, and max, defaulting to max, to dynamically control reasoning depth and balance response speed against inference quality.
• State management in multi-turn tool calls: In multi-turn tool-calling scenarios, maintaining conversational context requires passing back the reasoning_content and tool_calls returned by the model in the previous turn. This is necessary to prevent the model from losing track during long-running tasks.

6. Summary and Outlook

The release of this 2.8T-parameter model is not simply an increase in parameter count. It is a demonstration of systems engineering capability. From the mixed KDA/Gated DSA attention architecture, to the engineering trade-offs of MXFP4 quantization, to the focused optimization of agent tool-calling reliability, these developments point to a clear industry consensus: the competitive focus of large models is shifting from what they can do to whether they can do it well, reliably, and affordably.

For the developer community, this is not just a downloadable set of weights. It is one of the best available references for studying ultra-large sparse models. As inference engines continue to improve and hardware costs gradually decline, capabilities of this class are likely to move from leading research labs into production environments across industries, forming a solid foundation for the next generation of intelligent applications.