Hy4 Preview: Tencent's 770B Open Model and How to Adopt It

A preview model asks you to make a bet: adopt early and get the capability now, or wait for a final release that may behave differently. Hy4 preview makes that bet unusually easy to take, because Tencent released the weights under Apache 2.0 from day one. If the final model changes, the preview you tested does not disappear.

Tencent describes Hy4 preview as an early version of its next flagship, aimed at software engineering, office and data analysis, game development and scientific research. OpenRouter's catalogue dates it to 28 August 2026. This guide covers what is new, what Tencent's own numbers say against Hy3, and how to adopt a preview without being caught out.

What "preview" means here

Tencent is explicit about it. The model card calls Hy4 preview an early version with headroom left in both pre-training and post-training, and lists two known issues: the model can spend longer than necessary reasoning through complex tasks, and it tends to over-verify its own work. Tencent says it ships early on purpose, as it did with Hy3 preview, because outside feedback made Hy3 substantially better.

That history is worth reading as a forecast. Between Hy3 preview and the final Hy3, the architecture stayed the same, the vendor's benchmark figures moved by a few points, and the licence changed from a restrictive community licence to Apache 2.0. Hy4 preview already has the permissive licence, so the likely changes are in behaviour: reasoning length, tool-call stability and verification habits.

How to plan around a preview model covers the general playbook; the specifics for Hy4 follow.

A different generation of architecture

Hy4 preview is not a scaled-up Hy3. It is a mixture-of-experts model with 770B total parameters and 49B active per token, across 78 layers. The first layer is a dense feed-forward block; the remaining 77 are MoE layers with 256 routed experts plus one shared expert, routing each token to 8. A native multi-token-prediction layer (10B total, 0.7B active) supports speculative decoding.

Two changes stand out against Hy3. Attention is Gated DSA with IndexCache, a sparse-attention design, and the residual pathway uses what Tencent calls identity Hyper-Connections. Sparse attention is the design Tencent pairs with the jump in context: Hy4 preview supports 1M tokens, against Hy3's 256K.

The trade is cost per token. 49B active is more than double Hy3's 21B, which shows up directly in serving cost and in price. Mixture of experts explained covers the relationship between active parameters and what you pay.

Hy4 preview against Hy3, on Tencent's numbers

Both sets of figures come from the benchmark appendix on the Hy4 preview model card, run on Tencent's harnesses at each model's highest reasoning setting. Tencent notes that some Hy3 scores differ from those it published at the Hy3 launch because the harness, judge model and anti-hacking checks were updated, which is why Hy3's Terminal-Bench 2.1 figure here is 70.8 rather than 71.7.

• SWE-bench Pro — Hy3 57.9, Hy4 preview 65.7
• Terminal-Bench 2.1 — Hy3 70.8, Hy4 preview 85.4
• DeepSWE — Hy3 28.0, Hy4 preview 64.3
• APEX-Agents (pass@1) — Hy3 24.4, Hy4 preview 37.1
• GPQA Diamond — Hy3 90.9, Hy4 preview 92.3
• SWE-Marathon — Hy3 5.0, Hy4 preview 31.9

The pattern is clear. Knowledge and reasoning (GPQA) barely move. Long-horizon agentic work moves a lot: DeepSWE more than doubles, and SWE-Marathon goes from 5.0 to 31.9. The gains are concentrated in agent loops, not in quiz scores.

The same table places other models alongside. On Tencent's own Terminal-Bench 2.1 runs, Kimi K3 scored 85.7 and GLM 5.3 88.3, against Hy4 preview's 85.4; on SWE-bench Pro, GLM 5.3 scored 64.6 and Kimi K3 63.3, against 65.7. Kimi K3's own reported Terminal-Bench 2.1 figure is 88.3, a reminder that scaffolds move these numbers by several points. On Tencent's figures, Hy4 preview sits level with the leading open-weight models rather than clearly ahead of them. Terminal-Bench explained covers what that benchmark rewards.

Reasoning defaults to high, and that costs money

This is the migration trap. Hy3 defaults to no_think. Hy4 preview defaults to high. Swap the model name in a pipeline built for Hy3 and every request silently starts producing long chains of thought.

Combine that with the known over-reasoning issue and the price, and the cost can move a lot. OpenRouter lists Hy4 preview at $0.834 input, $2.501 output and $0.042 cached input per million tokens at the time of writing, against $0.132 and $0.528 for Hy3. Output is nearly five times dearer per token, and a high-effort preview may emit several times more output tokens for the same task.

Set reasoning_effort to no_think explicitly for anything that does not need deliberation, and measure output tokens per task, not just per-token price. The token cost of reasoning models shows how to budget for it.

1M context, 64K output

The 1M-token context matches the large-window group of open models. OpenRouter lists a maximum output of 64,000 tokens, which is enough for long code changes but lower than some peers; a single response cannot rewrite a very large file in one go. Plan agent steps around that ceiling.

Self-hosting is a serious deployment. Tencent's vLLM and SGLang examples use eight-way tensor parallelism, and an official FP8 checkpoint is published alongside the BF16 weights. Hosted access is also available through third-party inference providers.

Adopting a preview without getting burnt

1. Pin the exact weights or endpoint you evaluated. Apache 2.0 lets you keep a copy, so a changed final release cannot break you overnight.
2. Run your own long-horizon tasks at high and low effort, and record turns, output tokens and wall-clock time as well as pass rate.
3. Watch for the over-verification habit: repeated test runs and re-reads after the task is already done. Cap turns in your harness.
4. Keep Hy3 as the default for routine work and route only the long, hard agentic tasks to Hy4 preview, where its gains are largest.
5. Re-run the evaluation when the final Hy4 ships, and expect behaviour, not licence, to be what changes.