Hy4 preview

Spec

Attribute	Value
Developer	Tencent
Released	2026-08-28
Announced	2026-08-28
Context window	over 1,000,000
Pricing	$0.834/M input · $2.501/M output
License	Apache-2.0
Availability	Hugging Face (BF16 + FP8), ModelScope, GitCode, CNB, Tencent Cloud TokenHub, OpenRouter, WorkBuddy, CodeBuddy, Yuanbao, ima
Context window is recorded as the coverage states it — "exceeding 1M
tokens", also reported flatly as "a 1M-token context". No config.json value was
read, so the row is not given the exact integer a config file would supply; the
same figure for GLM-5.3 was only pinned once its config was read
(source).

Pricing is the API rate, and the weights are free. Apache 2.0 covers both the BF16 checkpoint and a separate FP8 quantisation, so the per-token figure prices Tencent's hosting, not the model.

Release Date

Released 2026-08-28, open-sourced the same day — announcement and weights together, with no staging window. That is the opposite of the two nearest comparisons: GLM-5.3 was announced 2026-08-14 and held its weights back about two weeks pending a safety evaluation, and Qwen3.8-Flash-Next ran a public countdown timer (source).

Benchmarks

Every figure below is Tencent's own, and no harness is named for any row. No third party has measured this model: it does not appear in the Artificial Analysis leaderboard captured the same morning, which lists Tencent only through the earlier Hy3 at Intelligence Index 42 (source).

Benchmark	Hy4 preview	Comparison as reported
Terminal Bench 2.1	85.4	reported as surpassing DeepSeek V4 Pro
DeepSWE	28.0 → 64.3	the 28.0 baseline model is not named
Internal blind evaluation (203 engineering tasks, 163 experts)	2.99 / 4	Kimi K3 2.94, GLM-5.3 2.92
Head-to-head vs GLM-5.3, same evaluation	46.8% wins · 12.8% ties · 40.4% losses	—
BenchAlign leaderboard	#7 of 228, 79.16/100	third-party leaderboard, read via BenchLM rather than from BenchAlign
**The blind evaluation was designed, run and scored by the party being
measured.** 163 of Tencent's own experts on 203 of Tencent's own tasks is a
useful internal signal and is not a comparison anyone else can reproduce. The
margin it reports — 2.99 against 2.94 and 2.92 — is **0.05 and 0.07 on a
4-point scale**, and nothing read gives a variance, a confidence interval or an
inter-rater agreement figure, so whether the ordering is separable is not
established.

"Never the lowest scorer in any single test" across 12 benchmarks is the claim Tencent leads with. Only the rows above were quoted in what was read; the other benchmarks are not named, so the claim cannot be checked here.

Use Cases

Agentic coding and terminal work, on the evidence of what was benchmarked: Terminal Bench, DeepSWE, and an internal task set described as engineering tasks. The first-party products it shipped into — CodeBuddy and WorkBuddy — point the same way (source).

The sparsity is the deployment argument. 49B active out of 770B is roughly 6% of the model per forward pass, which is what lets a 770B model be quoted at $0.834/M input. Nothing read measures throughput, latency or memory footprint, and no local-inference figure exists despite the model surfacing here through r/LocalLLaMA.

Compared To

• GLM-5.3 — the direct comparison Tencent chose, and the licences diverge more than the scores do. GLM-5.3 ships under a bespoke glm-5.3 licence requiring a Z.AI security review of any Model-as-a-Service operator above $10 billion group revenue; Hy4 preview ships Apache 2.0 with no such clause. On Tencent's own blind evaluation the gap is 2.99 vs 2.92.
• DeepSeek V4-Pro-0813 — 1.6T/49B MoE, MIT, the same 49B active parameters at roughly twice the total size. See ## Conflicting Reports for the Terminal Bench 2.1 figures, which do not line up.
• Kimi K3 — 2.8T MoE, the third model in Tencent's blind evaluation at 2.94.
• Qwen3.8-Flash-Next — the other open-weight Chinese release this fortnight, and the opposite design point: 176B total / 6B active against 770B/49B, qwen-community-1.0 against Apache 2.0.
• Architecturally it is closest to DeepSeek's work, by Tencent's own naming: the attention module is Gated DeepSeek Sparse Attention (Gated DSA). A lab publishing a frontier model built on a named competitor's attention mechanism is what open weights are supposed to produce, and it is worth recording as an instance rather than a generality.

Conflicting Reports

Terminal Bench 2.1 — Hy4 preview's 85.4 is reported as surpassing DeepSeek V4-Pro, and this wiki holds 87.9 for that model. DeepSeek V4-Pro-0813 carries a vendor-stated Terminal Bench 2.1 of 87.9 with no published harness, recorded here on 2026-08-14. Tencent's claim requires a V4-Pro figure below 85.4, and no figure Tencent used for V4-Pro was read.

This is not resolved and should not be. Two vendors self-reporting the same benchmark, each without naming a harness, is precisely the failure Eval Harness Configuration exists to track: the numbers may both be correct measurements of different configurations, and nothing published lets anyone tell. Per CLAUDE.md, the page body carries neither as settled and both sit here with their sources (source).

Parameter shape shared with GLM-5.3. The backbone reported here — 78 layers, 256 routed experts, top-8 active — is the same layer/expert/routing shape read from GLM-5.3's config.json on 2026-08-28. Recorded as an observation only: nothing read connects the two, and two MoE models arriving at the same configuration is not evidence of anything on its own.

Sources

• Tencent Hunyuan: Hy4 preview released and open-sourced, captured 2026-08-30
• Artificial Analysis leaderboard, 2026-08-30