MartechARTICLE

MetaX adapts C-series GPUs to Tencent's Hy4 on launch day and releases FlagOS W8A8 weights

MetaX adapts C-series GPUs to Tencent's Hy4 on launch day and releases FlagOS W8A8 weights
Image: Lucas Cardoso
Article · Martech
Reading
5 min
Tables
1
In short

MetaX says its C550 GPUs already support Tencent's flagship Hy4 model (Hunyuan), released on November 20, and published a quantized version via FlagOS. The manufacturer's own numbers show throughput gains and worse initial latency.

What MetaX announced

Chinese GPU maker MetaX says it completed a "day-zero" adaptation of its C-series GPUs to Hy4, Tencent's flagship model, hours after the company released the model's weights for download on November 20, 2026. Alongside this, MetaX published a W8A8 quantized version of Hy4 built on FlagOS, according to a Pandaily report published on November 22.

It's important to separate two things the news bundles together. One is the hardware and software adaptation to run the original Hy4 on MetaX's C550 GPUs. The other is this W8A8 version, which is not a new model trained from scratch: it's a lighter way of serving the same weights, using eight bits for weights and activations instead of the original precision. Pandaily also notes that this release is separate from MetaX's earlier adaptation to "Hy4 flash" (a leaner variant of the same model) and from Tencent's original announcement, covered in another Pandaily report.

MXMACA and FlagOS: the stack trying to play CUDA's role

MetaX attributes the quick integration to its MXMACA software stack, which includes drivers, interfaces, the MXCC compiler, operator adaptation, and connections to inference frameworks. In practice, it's an attempt to offer Chinese GPUs the same kind of layer that CUDA offers Nvidia GPUs: without it, running a large model on different hardware requires rewriting operator by operator.

FlagOS, behind the W8A8 version, is a BAAI (Beijing Academy of Artificial Intelligence) project to standardize operators across different Chinese chips. Pandaily published another piece on FlagOS version 2.2, which already covers 1,454 operators across manufacturers such as MetaX and Huawei Ascend. The stated goal is to let the same quantized model run on multiple domestic accelerators without manual rewriting, which matters in a market where access to cutting-edge Nvidia GPUs remains restricted by US export controls.

The 32-GPU benchmark: throughput rises, initial latency gets worse

The model card accompanying the release describes a cluster of four servers, each with eight 64GB C550 GPUs, totaling 32 GPUs. The test uses 64 concurrent requests, with roughly 4,000 input tokens and 1,000 output tokens per request, according to the report. These specific hardware and load conditions matter for interpreting the numbers, because they are not equivalent to a production test with real traffic.

MetricWith decode-graphWithout decode-graph
Aggregate output throughput80.98 tokens/s52.31 tokens/s
Average time to first token (with queue)227.2 s220.5 s

In other words: the "decode-graph" mode recommended by MetaX increases throughput by about 55%, but time to first token got worse, not better, when the waiting queue is counted in. Both numbers come from the same card and the same report, and together they rule out any reading that "everything improved": there is a real trade-off between aggregate throughput and latency perceived per individual request.

Two additional points limit the benchmark's scope. The deployment recipe uses a maximum context of 65,536 tokens, well below the 1 million token capacity Tencent advertises for the original Hy4, so the numbers say nothing about long-context performance. The card also marks two multimodal evaluations as "in progress," not completed.

In short: the numbers come from the manufacturer itself, specific to one configuration, and Pandaily was explicit in treating them as "vendor claims," not as independently verified results in real applications.

What Pandaily's verification found (and didn't find)

It's worth highlighting a reporting detail that sets this coverage apart from a simple reproduced press release. The cited FlagRelease repository is public and contains weight files in safetensors format available for download, not just an announcement or a hosted demo. In the verification, a limited request managed to retrieve 64 bytes from a weight fragment listed in the repository.

Pandaily did not download the complete model or run the deployment described in the card. This means the existence of the weights was confirmed, but the performance MetaX declared was not independently reproduced.

The license: why the Apache 2.0 tag doesn't tell the whole story

The most delicate point for anyone considering commercial use of these weights is licensing. The derived repository's metadata carries an Apache-2.0 tag, but Tencent's original license for Hy4, called Hy License 1.0, explicitly states that it is not Apache 2.0 and treats quantized weights as a derivative work subject to the same conditions.

Hy License 1.0 imposes a separate license requirement for businesses offering the model as Model-as-a-Service (MaaS) at large scale, with an exception for internal use under specific conditions. In practice, the FlagRelease repository's Apache-2.0 tag describes the adaptation's technical packaging, it does not replace the commercial terms that come from above, from Tencent. Anyone evaluating the quantized Hy4 for a commercial product needs to read the original Hy License 1.0, not just the distribution package's license.

What this means for those deciding AI infrastructure

For teams in Brazil that today run open models mostly on Nvidia GPUs with CUDA, via llama.cpp, Ollama, or TensorRT-LLM, the episode is less about which GPU to buy and more about a sign of maturation: Chinese GPUs like MetaX's C550, combined with stacks like MXMACA and FlagOS, are chasing day-zero adaptation parity with large models, something that until recently was almost exclusively the domain of the Nvidia ecosystem.

This matters mainly to those evaluating Chinese clouds or infrastructure providers outside the Nvidia axis for cost or availability reasons, including outside China. For everyone else, the practical takeaway is more immediate: before adopting a third-party quantized weight, check the license at the model's original source, not the tag of the repository redistributing the weights, and treat performance numbers published by a manufacturer as a starting point for your own testing, not as a final measurement.

Translated from the Brazilian Portuguese original · Read the original