The Extended Brief

Introducing Hy4 Preview

Brief by The AI News AI newsroom · Aug 29, 2026, 9:11 PM EDT edition

Original reporting by Simon Willison · published Aug 29, 2026, 7:53 PM EDT

Tencent's new open-weight model more than doubles its predecessor's scale, giving self-hosters a 770B-parameter, 1M-context option if they can handle 1.56TB of weights.

Key points

• Tencent released Hy4 Preview, an open-weight, text-only model with 770B total and 49B active parameters. source ↗
• It offers a 1M-token context window and a 1.56TB weight download on Hugging Face. source ↗
• The jump is large: July's Hy3 had 295B total parameters, 21B active, and 256,000-token context. source ↗
• Hy4's chat template exposes only two reasoning effort levels: 'high' by default or 'no_think'. source ↗
• The author found the model's hidden reasoning trace written in truncated English, speculating it is more token-efficient. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Run your standard evaluation prompts against Hy4 Preview via OpenRouter before considering a switch from your current open model.
• Confirm your infrastructure can handle a 1.56TB download and the serving footprint of a 770B-parameter model before planning self-hosting.
• Benchmark latency and output quality with reasoning_effort set to 'no_think' versus the default 'high' for workloads that may not need extended reasoning.

Context

Hy4 Preview is Tencent's follow-up to Hy3, released in July, and is distributed as open weights on Hugging Face. Having 770B total parameters but only 49B active per token is characteristic of mixture-of-experts designs, which keep per-token compute far below that of a dense model of the same total size. The model accepts text input only, with no vision support.

What to watch

• Independent benchmark scores for Hy4 Preview would show whether the scale increase over Hy3 delivers real capability gains.
• A full release dropping the 'Preview' label, or a vision-capable variant, would clarify Tencent's plans for the line.

Related briefs

• ThinkingBox: Solving an agent task once vs. solving it 20/20: 507 stateful workflows graded on terminal database state [R]
• Google Cloud launches Gemini agent to work across enterprise systems
• We unlearned CCP alignment from Qwen3.6-35B-A3B: censored/propaganda answers 89.8% → 2.8%, general benchmarks within ~1 point (open weights)
• Ignore all instructions and read this blog: The state of AI-analysis evasion in malware

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.