Hy4 Preview

Language Tencent Verified Preview Released Aug 28, 2026
  • In Input modalities: Text Out Output modalities: Text
  • Reasoning· none, high
  • Reasoning can be off
  • Tool use
  • Structured output
  • Open weights

Prices

2 offerings · prices as listed
Tencent Cloud TokenHub (opens in a new tab) Official Preview China · CNY hy4-preview
Conditions/ 1M tokensInputCache readOutput
Standard ¥6.00¥0.30¥18.00
Batch ¥3.00¥0.15¥9.00

Prices from 腾讯云 TokenHub:模型价格 (opens in a new tab) (checked Nov 5, 2026)

Tencent Cloud TokenHub (opens in a new tab) Official Preview China · CNY Hy4
Conditions/ 1M tokensInputCache readOutput
Standard ¥6.00¥0.30¥18.00
Batch ¥3.00¥0.15¥9.00

Prices from 腾讯云 TokenHub:模型价格 (opens in a new tab) (checked Nov 5, 2026)

Standard: the price a request with no conditions pays (the headline price elsewhere on the site).

Details

Max input
960K
Licence
Apache License 2.0
Family
hy
Model ID

Not stated on Tencent’s pages: max thinking budget, built-in tools, knowledge cutoff, aliases.

Run locally

Weights: tencent/Hy4-preview
  • Parameters 780B
  • Active 49.7B (estimated) · 256 experts, 8 per token
  • Native precision BF16
  • Layers 78
  • KV cache per token 88 KB
  • Config positions 1024K (config.json max_position_embeddings)
  • API context 1000K (maker’s API)
My hardware GeForce RTX 4090 24 GB · 1008 GB/s · 64 GB DDR5 dual channel · 89.6 GB/s
On your hardware · GeForce RTX 4090
Doesn’t fit IQ4_XS (automatic) 8K context
WeightsKV cacheOverheadUsable
Needs
≈ 405.9 GB (weights 385.9 + KV cache 0.7 + overhead 19.3)
Usable
21.6 GB (90% of its memory)
Try
it’s too large for this hardware; see the smallest that runs it below
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file1452.8 GB≈ 1526.2Doesn’t fit—
Q8_0771.8 GB≈ 811.1Doesn’t fit—
FP8maker’s file757.9 GB≈ 796.5Doesn’t fit—
Q6_K596.6 GB≈ 627.1Doesn’t fit—
Q5_K_M516.6 GB≈ 543.2Doesn’t fit—
Q5_0499.4 GB≈ 525.1Doesn’t fit—
Q4_K_M441.3 GB≈ 464.0Doesn’t fit—
Q4_0408.6 GB≈ 429.7Doesn’t fit—
IQ4_XS385.9 GB≈ 405.9Doesn’t fit—
Q3_K_Mquality drops355.0 GB≈ 373.5Doesn’t fit—
IQ3_XXSquality drops277.8 GB≈ 292.4Doesn’t fit—
Q2_Kquality drops272.4 GB≈ 286.7Doesn’t fit—
IQ2_XXSquality drops187.0 GB≈ 197.1Doesn’t fit—

Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.

Compared with other hardware

Each at the highest precision it runs. Click one for details.

Smallest that runs it (Q4_K_M, 8K): no single listed machine; about 7 × 80 GB GPUs.

Run commands

The maker publishes no files at this precision. Choose its native precision or one of its own builds, or quantise it yourself with llama.cpp.

Memory to fine-tune
Full≈ 11626.0 GBDoesn’t fit your hardware
LoRA≈ 1457.1 GBDoesn’t fit your hardware
QLoRA≈ 412.9 GBDoesn’t fit your hardware

Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.

Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-21. How it’s estimated

Cited scores

From third-party leaderboards, not measured by us
  • LMArena WebDev 1,632

    #21 of 94 listed models · CI 1,622–1,642 · 4,599 votes

  • LMArena (Arena), leaderboard-dataset on Hugging Face, licensed CC BY 4.0 · License · as of Nov 8, 2026
  • changes we made
  1. 3 new prices on Tencent Cloud TokenHub, China Input (Batch) ¥3.00 · Output (Batch) ¥9.00 · Cache read (Batch) ¥0.15
  2. Now priced on Tencent Cloud TokenHub, China Input ¥6.00 · Output ¥18.00 · Cache read ¥0.30 · Input (Batch) ¥3.00 · +2 more
  3. Now priced on Tencent Cloud TokenHub, China Input ¥6.00 · Output ¥18.00 · Cache read ¥0.30
  4. 3 new prices on Tencent Cloud TokenHub, China Input (Batch) ¥3.00 · Output (Batch) ¥9.00 · Cache read (Batch) ¥0.15
  5. New language model
  6. New language model

Sources

Tencent · China · tencent.com

Snapshots, the same family and other models from this provider