Tencent releases Hy4 preview, a 770B parameter MoE model. New Sherry quantization cuts size from 1.5TB to 214GB, enabling local inference on patched setups.

The barrier between frontier-scale artificial intelligence and local hardware just became significantly thinner. Tencent has released Hy4 preview, a massive 770-billion-parameter model designed for high-end cognitive tasks. While the raw scale of the model is impressive, the real disruption lies in how Tencent is making this behemoth accessible to those without a warehouse full of H100 GPUs.

The scale of Hy4 preview

Released on August 28, 2026, Hy4 preview is a Mixture-of-Experts (MoE) architecture. While it boasts 770 billion total parameters, it only utilizes 49 billion active parameters per token, allowing for a more efficient inference process than a dense model of the same size. One of its most potent features is a context window exceeding one million tokens, positioning it as a tool for deep-dive analysis of massive codebases or extensive scientific documentation.

Tencent has positioned Hy4 as a specialist in long-horizon work. The model is specifically tuned for software engineering, where it can plan, debug, and verify work across extended sessions. It also extends into game development, capable of building playable prototypes from a single prompt and refining them within game engines. In the realm of scientific research, Tencent claims the model can manage multiple Codex sessions in parallel, adjusting its research direction based on incoming results.

Breaking the 1.5TB hardware wall

For most enterprises and independent labs, a model with original BF16 weights totaling 1.5TB is effectively unusable locally. It requires an immense amount of VRAM that only the most expensive server clusters possess. This is where the AngelSlim team enters the frame. Using a specialized compression toolkit, Tencent has introduced GGUF builds that drastically reduce the footprint of the model.

The most aggressive version, the STQ1_0 build, shrinks the model to 213.66 GiB. This represents an 86% reduction in size. To put this in perspective, the team also provided a Q4_K_M build at 435.20 GiB and a UD-IQ1_M build at 219.83 GiB. By moving from 1.5TB to roughly 214GB, the model moves from the realm of data centers into the realm of high-end workstations.

How Sherry quantization works

The magic behind this reduction is a quantization method called Sherry. Rather than applying a uniform bit-reduction across the entire model, which typically destroys reasoning capabilities, Tencent employed a more surgical approach. The team used different quantization levels based on the sensitivity of each layer.

Non-critical layers are compressed as low as 1.31-bit, while sensitive layers retain 2-bit precision or higher. This results in an average of approximately 2.38 bits per weight (bpw). Specifically, Sherry utilizes 3:4 sparse ternary quantization: in every group of four weights, three are kept as ternary values and one is forced to zero. This is managed through a compact code and a table-selection bit, with one fp16 scale for every 256 weights.

The result is a model that retains a surprising amount of its original intelligence. On the SWE-bench Pro, the compressed version lost only 0.7% in performance. Other benchmarks showed a drop of only 0.2 to 1.6 points compared to the original BF16 version.

Local inference and the RTX 4090 reality

There is a common misconception that a 214GB model can simply be loaded onto a standard gaming laptop. It is important to clarify that while the files fit on a disk, full residency still requires roughly 214 GiB of VRAM for the STQ1_0 build. A single RTX 4090 laptop does not have 214GB of VRAM. However, the model can be run through patched local inference setups that were previously impossible.

Tencent tested heterogeneous device joint inference using prima.cpp. In one configuration, they combined an RTX 4090 laptop with a 4-A4000 server. Despite having only 80GB of total VRAM and 64GB of RAM, they achieved an inference speed of 1.02 tokens per second. This was roughly six times faster than attempting to run the model offloaded on the laptop alone. This proves that frontier-scale models can now be distributed across a mix of cheaper, mismatched hardware.

The useful news isn't that a normal laptop can run it. The useful news is that a 770-billion-parameter open model is starting to fit into stranger, cheaper hardware setups.

Performance benchmarks and accessibility

Tencent's internal testing suggests Hy4 is highly competitive. In a blind comparison where 163 experts rated outputs across 203 engineering tasks, Hy4 preview averaged 2.99, edging out GLM 5.3 (2.92) and Kimi K3 (2.94). Against GLM 5.3, it recorded 46.8% wins and 12.8% ties.

To ensure widespread adoption, Tencent released the model under an Apache 2.0 license. It is available via Tencent Cloud TokenHub, OpenRouter, and public repositories. For those preferring API access, the pricing is structured as follows:

• Cached input: .042 per million tokens
• Standard input: .834 per million tokens
• Output: .501 per million tokens

Global implications for AI enterprises

For businesses in the USA, UK, and other global markets, the release of Hy4 preview signals a shift in the economics of LLM deployment. The ability to run a 770B parameter model on heterogeneous, non-frontier hardware reduces the reliance on massive cloud expenditures for specialized R&D tasks.

In the US and UK, where data privacy and intellectual property are paramount, the move toward high-performance open-weight models allows companies to keep their most sensitive software engineering and financial modeling data on-premises. By using patched local inference, a firm can leverage a cluster of existing GPUs rather than investing in a dedicated AI supercomputer.

Furthermore, the Apache 2.0 license provides a permissive framework for commercial integration, allowing Western startups to build proprietary layers on top of Hy4 without the restrictive licensing often found in closed-source frontier models. While the model is an early preview and may occasionally over-verify its work or reason longer than necessary, its capacity to orchestrate complex tasks makes it a viable alternative for enterprises looking to diversify their AI stack away from a few dominant providers.

FAQ

---

Sources: Startupfortune, Lookonchain, Testingcatalog ·