Tencent

Tencent Hy4 preview

Tencent's open-weight 770B flagship under Apache 2.0, at $0.834/$2.501 per million tokens, flat across a 1M window. Two of seven tracked scores are independent, the rest are Tencent's own. That check, vals.ai's Terminal-Bench 2.1 run, measured 55.06, 30.3 points under Tencent's 85.4 on a different harness. Tencent's re-run Hy3 figures had sat within about a point of independent runs on three benchmarks.

Tencent Hy4 preview benchmarks and pricing, every number sourced: 7 tracked Tencent Hy4 preview benchmark scores (2 independently run, 5 still resting on a vendor’s own claim), priced at $0.83 per million input tokens and $2.50 per million output.

Tencent Hy4 preview architecture: Mixture-of-Experts; 770B total parameters (49B activated per token); 1M-token context window.

Released: 2026-08-28
License: open-weights
Context window: 1M tokens
Knowledge cutoff: Not disclosed
Verified: sources checked 2026-09-27–2026-10-08
Parameters: 770B (49B active)
Architecture: Mixture-of-Experts

Tencent Hy4 preview’s verified record

Tencent Hy4 preview’s most-compared rival is Qwen3.8-Max: 1 trail and 6 not callable across their 7 shared comparisons. Tencent Hy4 preview is priced at $0.83/$2.50 per 1M tokens (in/out) vs Qwen3.8-Max’s $2.00/$6.00.

Against the 188 head-to-head comparisons Tencent Hy4 preview shares with other tracked models:

14 real gaps2 inside the band172 not called

A gap counts for Tencent Hy4 preview only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Tencent Hy4 preview trails on 14 of them.

Terminal-Bench 2.1 Terminal ops

±10.6 is noise

Behind

MiMo-V2.6-Pro −12.7 · GPT-6 Luna −18.0 · Grok 4.7 −18.3 · DeepSeek V4.1 Flash −19.5 · MiMo-V2.6-Flash −21.3 · Kimi K3 −25.8 · Claude Sonnet 5.5 −28.1 · Claude Opus 5.5 −32.6 · 2 previous: GLM-5.2 −12.7 · GPT-6 Sol −28.1

Tie

1 model within ±10.6

Unverified

3 models — vendor-reported on one side

Setup-dependent

21 models — scored on a different harness or effort tier

Terminal-Bench 4.0 Terminal ops

±12.4 is noise

Behind

MiMo-V2.6-Flash −16.2 · MiMo-V2.6-Pro −23.2 · Step 5 Preview −23.7 · Qwen3.8-Max −26.3

Tie

1 model within ±12.4

Unverified

1 model — vendor-reported on one side

Setup-dependent

4 models — scored on a different harness or effort tier

No verdict for Tencent Hy4 preview anywhere on DeepSWE, HLE · no tools, HLE · with tools, Toolathlon-Verified ( nothing independently confirmed on both sides); GPQA Diamond ( saturated).

None of Tencent Hy4 preview’s reasoning comparisons are independently confirmed on both sides yet.
None of Tencent Hy4 preview’s coding comparisons are independently confirmed on both sides yet.

Tencent Hy4 preview API pricing

$0.83 in / $2.50 out per 1M tokens — official pricing source

What Tencent Hy4 preview costs per job

Tencent Hy4 preview cost for three reference workloads, computed from its list rates
Workload	Tokens in / out	Cost
One long chat turn	100K / 10K	$0.108
A codebase review	1,000K / 100K	$1.08
A day of agent work	10,000K / 1,000K	$10.84

Computed from Tencent Hy4 preview’s list rates above — cache discounts and batch tiers are not applied.

Tencent Hy4 preview benchmark scores

Tencent Hy4 preview benchmark scores, provenance, and source links
Benchmark	Score
Terminal-Bench 2.1^[1] Terminal ops · ±10.6 is noise	55.1vals.ai · 2026-10-01
DeepSWE^[2] Long-horizon coding · ±9.5 is noise	64.3huggingface.co · 2026-09-27
Toolathlon-Verified^[3] Multi-tool chores · ±9.7 is noise	74.1huggingface.co · 2026-09-27
GPQA Diamondsaturated^[4] Expert science Q&A — not ranked at any gap size	92.3huggingface.co · 2026-09-27
HLE(no tools)^[5] Reasoning · ±2 is noise	43.4huggingface.co · 2026-09-27
HLE(with tools)^[6] Reasoning · ±2 is noise	55.4huggingface.co · 2026-09-27
Terminal-Bench 4.0^[7] Terminal ops · ±12.4 is noise	8.1vals.ai · 2026-10-08

Who ran these numbers: 2 of 7 independent — vals.ai (2); vendor self-reported (5).

Terminal-Bench 2.1: vals.ai's own Terminus 2 run, board row 'Hy4 Preview' (55.06 ± 3.43, no effort parameter listed, pass@1), from its archived Terminal-Bench 2.1 table. Replaces Tencent's self-reported 85.4 (model-card appendix, Claude Code harness, up to 500 turns and a 12-hour timeout per trial), 30.3 points higher. Tencent's launch appendix reports 85.4 for this cell (Claude Code harness, up to 500 turns, highest reasoning) — a different setup from the independent run stored here, and not stored as a row.
DeepSWE: Tencent's model-card appendix figure at the highest reasoning setting, run on mini-swe-agent (8 CPUs, 16 GB per task) — the same harness the independent deepswe.datacurve.ai board uses, but Tencent's own run. That board did not list Hy4 on 2026-09-27.
Toolathlon-Verified: Tencent's model-card appendix figure at the highest reasoning setting, run on Tencent's internal agent scaffold with some MCP tools reimplemented and a 2-hour timeout instead of the official 5,400 seconds. toolathlon.xyz lists the predecessor Hy3 (56.8) but not Hy4 as of 2026-09-27.
GPQA Diamond: Tencent's model-card appendix figure at the highest reasoning setting; harness not stated. GPQA Diamond is graded saturated on this site. Artificial Analysis had no Hy4 entry on 2026-09-27.
HLE: Tencent's model-card appendix figure (text-only subset, no tools) at the highest reasoning setting. Artificial Analysis had no Hy4 entry on 2026-09-27; for the predecessor, Tencent's re-run figure (34.4) sat within a point of AA's independent run (33.5).
HLE: Tencent's model-card appendix figure (text-only subset, with tools) at the highest reasoning setting. No independent with-tools run found.
Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 8.081 at no effort tier stated. No official or Artificial Analysis Terminal-Bench 4.0 row exists for Tencent Hy4.

Notes on the record

Pricing per Tencent's announcement: $0.834 input / $2.501 output per million tokens, $0.042 cached. Tencent Cloud's TokenHub lists the model at ¥6 / ¥18 (¥0.3 cache hit) with no length condition and no peak/off-peak rate — one flat rate across the window (checked 2026-09-27). No API launch promotion.

Weights are on Hugging Face (tencent/Hy4-preview) under Apache 2.0: a 770B-parameter mixture-of-experts with 49B active, a 1M-token context (TokenHub caps input at 960K), 64K max output, text in and text out. Two reasoning modes, no_think and high (the default). Knowledge cutoff not disclosed. Tencent calls it an early version shipped with known issues, including "spending longer than necessary reasoning" and "a tendency to over-verify its own work."

Two of the seven tracked scores is independent; the rest come from the benchmark appendix on Tencent's model card, at the highest reasoning setting. The independent one, Terminal-Bench 2.1, is vals.ai's independent Terminus 2 run, 55.06, which on 2026-10-01 replaced Tencent's 85.4 from Claude Code with up to 500 turns and a 12-hour timeout. Other harnesses: Toolathlon-Verified on Tencent's own scaffold, with some MCP tools reimplemented and a longer timeout; DeepSWE on mini-swe-agent, the harness the independent board uses. Tencent also reports Agents' Last Exam on the 105-task ALE-CLI split (22.8); this site tracks the full split, so that figure is not recorded. No other tracked board had a Hy4 row on 2026-09-27.

The appendix includes Tencent's re-run of Hy3 — its notes say some Hy3 scores changed after harness, judge and anti-hacking updates — at 34.4 on HLE without tools, 90.9 on GPQA Diamond, 56.2 on Toolathlon-Verified and 70.8 on Terminal-Bench 2.1. Independent runs land within about a point on the first three — 33.5 and 89.7 (Artificial Analysis), 56.8 (Toolathlon's own board) — and 6.4 points lower on Terminal-Bench 2.1 (Artificial Analysis, 64.4), inside the 10.6-point band. Hy4's own independent run breaks that pattern: 30.3 points under Tencent's figure.

Compare with

Gemini 4 Hydrogen Gemini Decision 1 Muse Ember Claude Fable 5.2 any model…

FAQ

Has Tencent Hy4 preview been independently benchmarked?

On two benchmarks. vals.ai's archived Terminal-Bench 2.1 board lists Hy4 preview at 55.06, against Tencent's own 85.4, and that run now replaces the vendor figure here; vals.ai's Terminal-Bench 4.0 run (8.08, default effort) joined on 2026-10-08. Artificial Analysis, tbench.ai, Toolathlon, the DeepSWE board, ARC Prize, LiveBench, Snorkel's Agents' Last Exam and MathArena had no Hy4 row when checked on 2026-09-27, and vals.ai no longer runs new models on GPQA Diamond, SWE-bench Verified or LiveCodeBench. The other five scores are Tencent's own, read from its model card.

How much does Tencent Hy4 preview cost?

$0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042, per Tencent's announcement. Tencent Cloud's TokenHub lists ¥6 and ¥18 per million tokens with no length-based tier and no peak/off-peak pricing, so the rate is flat across the 1M-token window. It is also sold through OpenRouter at the same USD rates.

Is Tencent Hy4 preview open source?

The weights are. Tencent published Hy4 preview and an FP8 version on Hugging Face under the Apache 2.0 license on 2026-08-28, alongside hosted access through Tencent Cloud's TokenHub and OpenRouter. The model is text-only, with two reasoning modes: no_think and high.

How far can Tencent's own benchmark numbers be trusted?

Less than the predecessor suggested. Tencent's re-run Hy3 figures sit within about a point of independent results on Humanity's Last Exam, GPQA Diamond and Toolathlon-Verified, and 6.4 points above Artificial Analysis on Terminal-Bench 2.1. But the one independent check of Hy4 itself, vals.ai's Terminal-Bench 2.1 run, measured 55.06 against Tencent's 85.4, a 30.3-point gap. Part of it may be the harness: Tencent ran Claude Code with up to 500 turns and a 12-hour timeout, vals.ai ran Terminus 2. Treat the other five Tencent figures with that gap in mind.

How does Tencent Hy4 preview compare with Hy3?

On Tencent's own table the jump is large: DeepSWE from 28.0 to 64.3, Humanity's Last Exam without tools from 34.4 to 43.4, Terminal-Bench 2.1 from 70.8 to 85.4. vals.ai's independent Terminal-Bench 2.1 run of Hy4 reads 55.06, below Tencent's own Hy3 figure. Tencent calls it the largest generation-over-generation gain it has measured. Hy3 is still sold on TokenHub at a lower price, and no Tencent page describes Hy4 as replacing it.

What is Tencent Hy4 preview's best independent score?

Terminal-Bench 2.1 at 55.06 (2026-10-01). The independent side of this record remains thin — Tencent's launch figures rest on the company's own presentation materials rather than third-party runs, so the verified column is short while the preview settles.