Tencent releases Hy4 Preview open-weight text LLM with 770B params and 1M context
Tencent has released Hy4 Preview, an open-weight text-only large language model that substantially expands the footprint of its prior Hy family release. The model carries 770 billion total parameters with 49 billion active during inference, supports a one-million-token context window, and weighs in at 1.56 terabytes on Hugging Face. Compared with July's Hy3, which offered 295 billion total parameters, 21 billion active, 256,000 tokens of context, and a 598-gigabyte package, the new preview represents a sharp scaling step across every advertised dimension. For builders evaluating long-context open models, the release signals Tencent's continued push into frontier-scale public weights. Simon Willison's coverage highlights the magnitude of the jump, though the packet does not include independent benchmarks or licensing details beyond the open-weight framing.
Tencent releases Hy4 Preview open-weight text LLM with 770B params and 1M context
Tencent shipped Hy4 Preview, an open-weight text-only LLM with 770B total parameters, 49B active parameters, a 1M token context window, and a 1.56TB weight package on Hugging Face. Simon Willison reports a large jump from July's Hy3, which listed 295B total parameters, 21B active, 256,000 context, and 598GB.
Key takeaway
Tencent's Hy4 Preview pushes open-weight text LLMs to 770B total / 49B active params and a 1M-token context, roughly tripling Hy3's scale from July.
What happened
Tencent released Hy4 Preview, an open-weight text-only LLM with 770 billion total parameters and 49 billion active parameters, according to Simon Willison's reporting on August 29, 2026.
The model supports a one-million-token context window and is distributed as a 1.56 terabyte package on Hugging Face, up sharply from July's Hy3 release at 295 billion total parameters, 21 billion active, 256,000 context, and 598 gigabytes.
Evidence
Tencent released Hy4 Preview as an open-weight text-only LLM with 770B total and 49B active parameters
Simon Willison · attributed
New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, 1.56TB on Hugging Face
Hy4 Preview supports a 1M token context window and ships as a 1.56TB package on Hugging Face
Simon Willison · attributed
770B total parameters, 49B active parameters, 1M token context window, 1.56TB on Hugging Face
Hy4 Preview is a large size increase over July's Hy3 across parameters, context, and weight package
Simon Willison · attributed
This is a big size increase from their previous Hy3 in July, which was 295B, 21B active, 256,000 context, 598GB
Why it matters
Operators weighing self-hosted long-context stacks must now budget for 1.56TB model weights and verify whether their hardware can serve a 49B-active-parameter model at production latency.
Limits and uncertainties
The evidence packet does not include independent benchmark results or production serving guidance for Hy4 Preview.
Licensing terms beyond the open-weight framing are not specified in the available reporting.
Simon Willison's article excerpt in the packet is truncated and does not include the full post text.
Practical implications
Teams evaluating self-hosted open models need storage and download capacity for a 1.56TB Hugging Face weight package.
Infrastructure planners should compare Hy4's 1M-token context and 49B active parameters against current Hy3 deployments at 256,000 context and 21B active.
What to watch
Whether Tencent or third parties publish benchmark scores and serving requirements for Hy4 Preview.
Hugging Face model card updates clarifying license terms and recommended inference setup for the 1.56TB release.
