Key details

  1. Tencent released Hy4 preview on August 28, 2026.
  2. The backbone has 770B total parameters and 49B activated parameters per token.
  3. The model has 78 backbone layers and a one-million-token context window.
  4. A native MTP layer is included for speculative decoding.
  5. Full and FP8 model weights are published under Apache 2.0.
  6. The Hugging Face full-weight repository is approximately 1.56TB.
  7. Tencent provides vLLM and SGLang deployment recipes plus fine-tuning and quantization tooling.
  8. Tencent’s internal blind evaluation involved 163 experts and 203 engineering tasks; the reported advantage over GLM-5.3 and Kimi K3 is vendor-produced.

What builders should take away

  1. Benchmark Hy4 on your own coding, retrieval and tool-use traces before comparing it with hosted frontier models; the public weights make direct evaluation possible.
  2. Plan hardware from the checkpoint footprint and KV-cache requirements rather than the 49B active-parameter number alone.
  3. Use the FP8 checkpoint and official serving recipes as the first production experiment instead of converting the full checkpoint yourself unless you have a specific reason.
  4. Measure long-context accuracy and latency at the context lengths your application actually uses; a 1M maximum is not evidence that every workload benefits from filling it.
  5. Keep tool execution permissioned and observable while testing the model’s agent features, especially because this is an early preview with documented over-reasoning behavior.

What changed

On August 28, 2026, Tencent released Hy4 preview as open weights on Hugging Face and other model hubs. The Mixture-of-Experts model has 770B total parameters with 49B activated per token, 78 backbone layers, a one-million-token context window and a native MTP speculative-decoding layer. Tencent publishes both full and FP8 checkpoints under Apache 2.0 and provides vLLM and SGLang serving recipes, tool-call and reasoning parsers, fine-tuning guidance and quantization tooling. Reuters independently confirmed the release and Tencent’s positioning around software engineering, research and analytical work.

Why it matters

Hy4 adds another frontier-scale open model that builders can actually inspect and run behind their own infrastructure rather than accessing only through a proprietary API. The 49B-active sparse design can reduce per-token compute relative to its 770B total size, but the practical deployment floor remains high: the full Hugging Face repository is about 1.56TB and Tencent’s published FP8 serving recipe uses eight-way tensor parallelism. For organizations that can operate that class of hardware, the release expands model-control and privacy options for long-context coding and knowledge work. Smaller builders are more likely to consume it through hosted inference or later quantizations.

The release is open enough to benchmark independently

Tencent publishes the model under Apache 2.0 and provides full and FP8 checkpoints plus source recipes. That means the launch can be evaluated on private repositories, long documents, agent tools and organization-specific tasks rather than relying only on Tencent’s product demos.

Sparse activation does not make this a desktop model

Only 49B of 770B backbone parameters are active per token, but all model weights still need to live somewhere. The full Hugging Face repository is about 1.56TB and the FP8 variant remains hundreds of gigabytes. Tencent’s vLLM example uses tensor parallelism across eight GPUs, so builders should distinguish active-compute efficiency from memory footprint.

Tencent is optimizing for long-horizon productivity work

Tencent describes Hy4 as stronger in software engineering, financial and office analysis, game development and scientific research. The architecture provides a 1M-token context window and native speculative decoding, while the serving stack includes automatic tool choice and model-specific tool/reasoning parsers.

The quality evidence is useful but vendor-led

Tencent reports a blind internal evaluation in which 163 employees rated 203 engineering tasks and Hy4 scored slightly ahead of GLM-5.3 and Kimi K3. That is more concrete than a marketing adjective, but it remains an internal Tencent evaluation with company-selected tasks and raters. Independent public harnesses will be more informative for cross-model decisions.

Tencent itself labels this an early model

The project documents known tendencies to over-reason and over-verify and calls Hy4 an early preview with further pre- and post-training work planned. Teams should expect model and serving behavior to evolve rather than treating the preview as a frozen production baseline.

What to watch next

  • Independent coding, agent and long-context evaluations using public harnesses.
  • Hosted-provider pricing and availability for Hy4 outside Tencent Cloud.
  • Smaller or more aggressively quantized Hy4 variants that reduce the current server-scale memory requirement.
  • Whether Tencent addresses over-reasoning and over-verification in the next Hy4 checkpoint.
  • Measured throughput and total cost on common H100/H200/B200-class deployments.

Still unclear

  • Tencent’s strongest launch comparisons are internal and have not yet been independently reproduced.
  • The practical serving cost depends heavily on quantization, context length, tensor parallelism, batching and hardware.
  • Hy4 is explicitly labeled a preview, so checkpoints, recommended parameters and serving behavior may change quickly.

Sources

Direct reading behind this dossier.

3 sources
Hy4-preview repository
Tencent Hy Team / GitHub primary

Primary architecture, benchmark, deployment, licensing and known-limitations documentation.

Hy4-preview model repository
Tencent / Hugging Face primary model repository

Public Apache-2.0 model files and repository footprint.

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment