MetaX adds Hy4 support on C series GPUs using FlagOS W8A8 weights
Read more
Pandaily
pandaily.com

MetaX adds Hy4 support on C series GPUs using FlagOS W8A8 weights

MetaX announced that its C series GPUs have completed zero-day adaptation for the flagship Tencent Hy4 model following the release of downloadable weights in November. Furthermore, the chip developer presented a FlagOS-based W8A8 version for its hardware, which is a deployment variant rather than the creation of a new, specially trained model.

The integration was achieved through MetaX's MXMACA software stack, which includes drivers, interfaces, the MXCC compiler, operator adaptation, and connections to inference frameworks. The announcement highlights high throughput, low latency, and stable operation, although these claims are vendor statements and not independently verified results across all applications.

The public FlagRelease repository contains safetensors weight files that can be downloaded, not just an announcement or demonstration. The W8A8 designation indicates eight-bit weights and activations. Upon verification, 64 bytes were successfully retrieved from the listed weight fragment; Pandaily did not download the full model or run the deployment.

The accompanying model card describes four servers, each equipped with eight MetaX C550 GPUs with 64 GB capacity, totaling 32 GPUs. In a test scenario with random input, 64 concurrent requests were used, each with approximately 4000 input and 1000 output tokens per request. These operating conditions and hardware specifications are important when interpreting the data obtained.

Within this test, the recommended graph decoding mode provided an aggregated output speed of 80.98 tokens per second compared to 52.31 tokens per second without it. However, the average time to the first token increased from 220.5 to 227.2 seconds, including queue waiting time. Consequently, the results do not confirm the general claim of improving all latency metrics.

The same card still notes two multimodal tests as ongoing. Its deployment recipe uses a maximum context setting of 65,536 tokens, which should not be confused with the claimed Hy4 context capacity of 1 million tokens. Performance metrics and evaluations remain data provided by the publisher and depend on the configuration.

Attention should also be paid to licensing. Although the production repository metadata contains an Apache-2.0 tag, the original Tencent Hy License 1.0 explicitly states that it is not Apache 2.0 and includes quantized weights as derivative works. It sets separate licensing requirements for qualified enterprises providing the Model as a Service, with an exception for internal use under certain conditions.

A useful addition is MetaX's specific service path and the availability of quantized weights. This differs from MetaX's earlier Hy4 adaptation and Tencent's initial model announcement; neither the hardware integration nor the repository metadata negates the commercial terms of the original model.

Similar stories

Tencent releases weights of flagship Hy4 model under conditional license
Read more
pandaily.com

Tencent releases weights of flagship Hy4 model under conditional license

Tencent has published the downloadable weights of its flagship Hy4 model on November 20, accompanying the release with a new System Card. This release followed the service launch on November 4 and differs from earlier Hy4 checkpoints that were available this week.

The official model card describes a multimodal 'mixture-of-experts' system with a language core containing 770 billion parameters. When activated, 49 billion parameters are used for each text token, and the context window reaches 1,048,576 tokens. The core count does not include the vision encoder, adapter, or DFly2 drafting component.

Tencent provided developers with various starting points for further training or deployment, including the Hy4 instructions, Hy4-FP8, and base Hy4 checkpoints.

Pandaily confirmed receiving a small range of bytes from one of the weight shards in each of the three repositories. However, the company did not download the full checkpoints nor conduct inference.

The deployment documentation contains recipes for vLLM and SGLang using eight-way tensor parallelism; these are significant infrastructure deployments, not proof that the flagship model runs on a standard laptop.

The weights use the Tencent Hy License 1.0, not Apache 2.0. Additional commercial terms apply to operators providing general-purpose remote inference or fine-tuning to third parties. Model-as-a-service operators must obtain separate written permission if their aggregate global gross revenue exceeds $50 million in any consecutive 12-month period, regardless of whether Hy4 generated that revenue, or before using the model in a product or service with more than 100 million active users in any calendar month.

The license excludes end-user workflows with limited functionality and simple relays to independently managed model services from its definition of model-as-a-service. A narrow exception for internal use requires that the model, its derivatives, or outputs are not provided to third parties. Therefore, these thresholds should not be presented as an all-encompassing restriction for every commercial user.

Tencent also notes the filtering of malicious training material for cybersecurity and the conducting of internal security audits by specialists. These assessments come from the issuer and are not independent guarantees that the model cannot be misused. Developers evaluating deployment must still conduct their own assessment of task performance, resource requirements, and security measures.

The practical change lies in access to the flagship model artifacts and the terms governing its use, rather than in a new debut of previously announced Hy4 capabilities in vision and computer usage.

Tencent releases Hy4 flash model weights for download in three variants
Read more
pandaily.com

Tencent releases Hy4 flash model weights for download in three variants

Tencent has published the downloadable weights for the Hy4 flash model following the announcement of its API on November 16th. This was reported by IT Home on November 18th. This update provides deployable checkpoints, rather than a re-presentation of the API specifications and pricing previously covered by Pandaily.

The official Tencent model page lists three versions: Hy4 flash, which is an instruction-tuned model; Hy4 flash-FP8, its instruction-tuned quantized version; and Hy4 flash-base, a pre-trained base model. The base checkpoint is intended for further training, not to replace the chat-ready variants.

Pandaily verified all three official repositories on Hugging Face. The weight index files referenced available shard lists, and anonymous byte range requests returned weight file data for each version. These checks confirm the availability for download but do not constitute full checkpoint downloads or successful output tests.

The standard checkpoint contains approximately 440.5 GB of tensor data, while the FP8 version is about 287.1 GB. These sizes refer to the weight data, not the minimum GPU memory requirements. Quantization reduces storage size but does not turn the release into a model suitable for small consumer devices.

Deployment documentation is an important part of this release. Tencent provides specialized vLLM and SGLang container images, including examples of eight-way tensor parallelism. Instructions for Transformers point to requesting model support rather than assuming that the standard released package already supports this architecture.

The vLLM recipe also requires host RAM, as Engram tables are defaulted to be offloaded to CPU memory. The initial configuration uses a context size of 8192 tokens; the model's claimed capacity of 1 million tokens does not guarantee that every hardware configuration can handle such length. The published memory hints are estimates, not measured minimum requirements.

The checkpoints use the Tencent Hy 1.0 license, which explicitly states that it is not Apache 2.0, despite being derived from it. Section 10 requires a separate written license for Model-as-a-Service operators whose cumulative global revenue exceeds $50 million over a consecutive 12-month period, or for using the model in a product or service with more than 100 million monthly active users. A narrowly defined exception applies to internal use when the model and its outputs are not provided to third parties.

Thus, this release is best described as open in terms of weights according to Tencent's stated terms, rather than as unrestricted open source. Tencent reports that the release of the larger Hy4 model weights is still in preparation.

TokenRhythm releases limited beta version of agent-oriented NeoHorse-1.5-Flash Preview model
Read more
pandaily.com

TokenRhythm releases limited beta version of agent-oriented NeoHorse-1.5-Flash Preview model

TokenRhythm, a Chinese artificial intelligence infrastructure developer managing a multi-model routing platform, unveiled NeoHorse-1.5-Flash Preview on November 5th. The company positions this model as agent-oriented, designed for task execution and software development. Access is exclusively available through a limited beta version, requiring users to apply via the company's official WeChat account.

According to TokenRhythm, the model was trained based on Qwen3.8-Flash and has a total of 180 billion parameters, with 6 billion actively utilized. In its self-comparison against the base model, the company noted an increase of 9.23 points in the Agents' Last Exam test and 5.92 points in the Natural Language to Repository test. Furthermore, improvements exceeding 20 points were achieved in physical science tasks, alongside progress in developer tooling, testing, web development, and cryptography and security. It should be noted that these evaluations were conducted by the company itself, and independent results have not yet been published.

This release is considered the second stage of TokenRhythm's work on recursive self-improvement. After releasing NeoHorse-1 in September, the company returned this model to its pool, gathered feedback on real-world task performance, and used it to train the next version. The company's routing system provides signals regarding required capabilities for tasks, and running several advanced models on the same task yields comparable results. Differences between the results of NeoHorse and these data are used as training signals. The team named three methods underlying this process: multi-model differential diagnosis, adversarial environment evolution, and online swarm learning.

This cycle also relies on OpenSquilla, an open-source product from TokenRhythm. In an update released the day after selecting the model level based on task intent, the selected level is maintained when continuing the task and increased upon retries.

TokenRhythm's public pages on GitHub and Hugging Face confirm the existence of the previous generation, not the new preview version. The NeoHorse-1 model was released on September 7th in 4B and 9B versions, fine-tuned on Qwen3.5 under the Apache 2.0 license, along with a technical report and quantized versions. At the end of September, the 4B decision-making model, named NeoHorse-Jev, followed. As of the publication, downloadable weights, a model card, or a license for NeoHorse-1.5-Flash Preview were absent from these pages, meaning it cannot be considered an open-weights release.

Popular