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.

Artefact size: 1.5 TB. Simon Willison gives 1.56 TB on Hugging Face (2026-08-29); Tencent's own quantisation announcement the same day says 1.5TB. These are the same quantity to different precision and are not recorded as a conflict (source) (source).

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).

Still true on 2026-08-31, checked rather than assumed. The 08-31 capture lists Tencent through Hy3 alone, unchanged at Intelligence Index 42, 256k context. Nothing in that capture moved on capability at all: every row's Intelligence Index and Cost per Task is byte-identical to 08-30, and only the three serving columns — Median Tokens/s, Latency First Chunk, Total Response — differ. Four days after release, the only numbers anyone has for this model are still the vendor's (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.

LMArena, 2026-09-13 — first appearance in this repo's top-10 series

Hy4 preview enters at #10 with 5.23% ±1.31%, listed under Tencent (source). It was absent from the 2026-08-30 and 2026-09-06 captures, and it takes the slot GLM 5.2 (Max) held on 09-06.

The metric is the leaderboard's own percentage with a confidence interval, not an Elo rating, and the snapshot sees only the server-rendered top 10 — so an absence in earlier captures is indistinguishable between too-few-votes, not-yet-listed and outside-the-top-ten, and none is asserted.

Two Chinese-lab entries are visible in the ten — this one and Kimi K3 at #8 — where GLM 5.2 held one of the two on 09-06. The count is unchanged; the occupants are not.

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 or latency.

2026-08-29 — the vendor shipped a ~200 GiB quantisation, one day after the weights, and said nothing measurable about what it costs. Tencent Hunyuan announced MIX-STQ1_0, compressing the model from 1.5 TB to ~200 GiB GGUF via its AngelSlim toolkit. The method is per-layer mixed precision chosen by calibration data rather than a uniform bit-width — some layers down to 1.31-bit (STQ1_0), some up to 2.06-bit (IQ2_XXS) — released as AngelSlim/Hy4-preview-GGUF (source).

The quality claim is "it still works well" and that is the whole of it. No benchmark table, no per-task delta and no retained-accuracy figure was published in what was read. A circulating "about 98% performance" is a Reddit submitter's headline, not Tencent's number and not traced to any table, and it is deliberately not carried onto this page. A 7.5× reduction is a real engineering result and an unquantified quality claim beside it is not a measurement; both halves are recorded.

This closes the gap this section named on 2026-08-30 only halfway. A local artefact now exists where none did, but ~200 GiB is still a multi-GPU or large-unified-memory figure, and no hardware target, throughput or memory measurement was published, so what it runs on remains unstated by anyone.

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.
• Hy3, its own predecessor — the one comparison where every figure is like-for-like. Simon Willison gives Hy3 (July 2026) as 295B total / 21B active / 256,000 context / 598 GB, against Hy4 preview's 770B / 49B / 1M / 1.56 TB. Total parameters and active parameters both roughly 2.3× in about seven weeks, context ~4×. This wiki holds no models/hy3 page, so the figures live here rather than on one; Hy3 is the only Tencent model Artificial Analysis lists, at Intelligence Index 42 (source).
• 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
• Tencent Hunyuan on the MIX-STQ1_0 GGUF quantisation, 2026-08-29
• Simon Willison, "Introducing Hy4 Preview", 2026-08-29
• Artificial Analysis leaderboard, 2026-08-31