Oversized Open-Weight Models: Licensing, VRAM, and Deployment Assessment

    This article assesses a newly released 2.8T-parameter open-weight model with roughly 1.56TB of weights, covering its MoE architecture, multimodal and sparse-attention features, MXFP4 quantization, benchmark caveats, custom license obligations, VRAM sizing, and a practical deployment checklist for enterprises.

    Start With the Conclusions

    The released weights are about 1.56TB, with 2.8T total parameters. For researchers, this provides a complete asset that can be inspected, evaluated, and fine-tuned offline. For enterprises, three things need to be confirmed before deployment: whether the license terms apply to the organization’s business scale, whether VRAM is sufficient to hold the full weights, and whether the inference framework supports the model’s quantization format.

    After all three checks pass, run a small-scale validation with your own workload data, then decide on the deployment method.

    Event Background

    The vendor has released the complete weights of its flagship model, as previously promised. Previously, this model was available only through an API and a web interface. Developers can now download, inspect, and deploy it directly. The weights are hosted in a public model repository, and no approval is required to obtain them. The repository contains 112 files, 96 of which are safetensors shards totaling about 1.56TB.

    According to the original article, this is the largest publicly downloadable open-weight model at the time of writing. That claim is based on the author’s search scope; readers should verify the current public model list before citing it.

    The direct value of such a model lies in three points: it enables local research on model behavior; it enables offline evaluation without relying on online services; and it enables customized fine-tuning based on the weights.

    Architecture Highlights

    1. Mixture-of-Experts (MoE) structure

    The vendor disclosed that the model activates only about 104B parameters per token. The model has 896 experts, each token is routed to 16 experts, and there are 2 shared experts. The model has 93 layers and a context length of 1 million tokens.

    This design reduces the computation per forward inference, but it does not reduce the storage required for the weights. During inference, all expert weights still need to remain resident in accessible VRAM or memory. Therefore, the main bottleneck is VRAM capacity, not compute.

    2. Multimodal capability

    The model includes an independent vision encoder and can directly process mixed text-and-image input. The vendor states that the model was specifically trained for agent scenarios and was continuously trained from the previous-generation base model.

    3. Sparse attention

    According to the vendor, compared with the base version, the model replaces the original 24 full-attention layers with a sparse attention mechanism. This change mainly reduces computation and KV cache usage in long-context scenarios. For tasks such as ultra-long documents and repository-level code context, this is a feature that deserves close attention.

    4. Quantization format

    The weights use the MXFP4 format, and activations use the MXFP8 format. These are products of quantization-aware training.

    In the repository file list retrieved for this article at the time of writing, no BF16 or FP8 versions were visible. This conclusion may change as the repository is updated; check the current repository before planning a project.

    Two practical implications need attention. First, the inference framework must natively support MXFP4; otherwise, the weights cannot be loaded. Second, whether the weights can be dequantized back to high precision and then requantized depends on the framework implementation and weight format, and it must be tested in practice.

    How to Read the Performance Data

    The benchmark results published by the vendor show clear differences in the model’s performance:

    • It performs strongly on terminal-based coding benchmarks and agent tasks, with some multi-step code tasks scoring higher than some leading closed models.
    • It is relatively weaker on scientific reasoning and high-difficulty knowledge benchmarks, with some scores clearly behind comparable closed models.
    • A third-party evaluation organization’s intelligence index ranked it fourth globally.

    There are three caveats for these data. First, vendor-reported data may contain selection bias; test sets, prompts, and tool-calling settings can all affect results. Second, benchmark scores are not equivalent to real business performance. Third, third-party rankings change over time, so the evaluation date and methodology should be cited.

    The safest approach is to run offline evaluations with your own real workload data. This is precisely the core advantage of an open-weight model over an API service: the evaluation process is controllable, and data does not leave the local environment.

    License: Open Weights Are Not Open Source

    The model uses a custom license agreement, not a standard open-source license such as MIT or Apache. The base terms of the agreement are close to the MIT style, allowing free use, copying, modification, distribution, fine-tuning, and commercial use.

    On that basis, the agreement adds two restrictions for large-scale commercial use:

    1. Companies that provide the model’s capabilities externally as a model-as-a-service and have cumulative revenue over 12 months exceeding a certain threshold need to sign a separate agreement before commercial use.
    2. When a commercial product’s monthly active users exceed a specified scale, additional terms are also triggered.

    The specific revenue threshold and monthly active user threshold should be based on the original license text. This article does not list specific numbers, to avoid compliance risk caused by misstatement.

    The table below outlines the approximate obligations for different types of users. Enterprises should rely on legal review results.

    User type License obligations Recommended action
    Individual learning and research Base terms, usually no additional obligations Use directly and retain the license file
    Small-team internal use Base terms, usually no additional obligations Confirm that internal use does not involve external services
    Commercial product below thresholds Use under base terms; monitor scale continuously Record user scale and revenue data
    Enterprise or SaaS service above thresholds Separate agreement required Legal review and signing before project approval

    Enterprises should especially note that business growth may trigger additional obligations at some future point.

    VRAM Estimate: Are 8 GPUs Enough?

    The deployment threshold is mainly determined by the size of the weights. During inference, VRAM must hold the following:

    Total VRAM requirement ≈ weight size + KV cache + activations + communication buffers + system reserve

    The weight size is about 1.56TB. KV cache grows with context length and concurrency, and it increases significantly in million-token context scenarios. Activations and communication buffers also require reserved space.

    Take an 8-GPU configuration as an example: if each GPU has 192GB of VRAM, the total for 8 GPUs is about 1.536TB, slightly below the size of the weights themselves. It cannot hold the full weights, let alone the KV cache. To reach about 2TB of total VRAM, each GPU would need at least 256GB.

    If per-GPU VRAM does not meet the requirement, one of the following measures can be taken: add more GPUs, switch to multi-node deployment, or use CPU or SSD offloading. Offloading significantly reduces inference speed, while multi-node deployment increases network costs and operational complexity.

    Deployment Checklist

    1. Hardware: total VRAM of at least 2TB; for an 8-GPU configuration, about 256GB or more per GPU. Prefer NVLink or equivalent high-speed interconnect.
    2. Storage: reserve at least 2TB of available space; NVMe SSDs are recommended.
    3. Download and verification: use a tool that supports resumable downloads to download shards. After downloading, use the sha256 manifest provided by the repository to run sha256sum -c and verify each shard.
    4. Inference framework: confirm that the framework version supports MXFP4 weight loading, MoE expert routing, and sparse attention. Record the version number for reproducibility.
    5. Parallel configuration: set the combination of tensor parallelism and expert parallelism according to the framework documentation. First validate output correctness with small batch requests.
    6. Service deployment: deploy through a container platform, monitor throughput, latency, and VRAM usage, and set alerts.

    For teams that do not have the above conditions, managed cloud inference services or rented GPU instances for elastic deployment are options. When evaluating cloud options, calculate costs as follows: GPU price per unit time × number of GPUs × runtime, plus storage and traffic fees. Throughput and latency should be measured with a fixed test prompt set for comparison. When using cloud services, also confirm the data storage location and check whether cross-border data transfer is involved.

    Conclusion

    This model allows the community to inspect, evaluate, and fine-tune a very large open-weight model, and its value is clear. However, its license terms, VRAM threshold, and benchmark performance all need separate evaluation. The recommended sequence is to first confirm whether the license applies, then calculate VRAM and framework support, then run small-scale validation with your own data, and finally decide on the deployment strategy.