Hy4 Flash Base

Language Tencent Partly verified
  • In Input modalities: TextImage
  • Open weights

Prices

0 offerings · prices as listed

No priced offering is listed for this model yet.

Details

Licence
tencent-hy-license
Family
hy-flash-base

Run locally

Weights: tencent/Hy4-flash-base
  • Parameters 219.4B
  • Active 12B · 512 experts, 16 per token
  • Native precision BF16
  • Layers 42
  • KV cache per token 47 KB
  • Config positions 1024K (config.json max_position_embeddings)
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
≈ 114.4 GB (weights 108.6 + KV cache 0.4 + overhead 5.4)
Usable
21.6 GB (90% of its memory)
Try
IQ2_XXS runs (needs system memory), but below 4 bits quality drops noticeably
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file408.7 GB≈ 429.5Doesn’t fit—
Q8_0217.1 GB≈ 228.3Doesn’t fit—
FP8204.3 GB≈ 214.9Doesn’t fit—
Q6_K167.8 GB≈ 176.6Doesn’t fit—
Q5_K_M145.3 GB≈ 153.0Doesn’t fit—
Q5_0140.5 GB≈ 147.9Doesn’t fit—
Q4_K_M124.1 GB≈ 130.7Doesn’t fit—
Q4_0114.9 GB≈ 121.1Doesn’t fit—
IQ4_XS108.6 GB≈ 114.4Doesn’t fit—
Q3_K_Mquality drops99.9 GB≈ 105.2Doesn’t fit—
IQ3_XXSquality drops78.2 GB≈ 82.4Doesn’t fit—
Q2_Kquality drops76.6 GB≈ 80.8Doesn’t fit—
IQ2_XXSquality drops52.6 GB≈ 55.6Needs system memory 24–36 tok/s—

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): GPU NVIDIA B200 (HGX B200); unified memory Apple M5 Ultra (64-core GPU) · 256 GB.

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≈ 3271.5 GBDoesn’t fit your hardware
LoRA≈ 411.0 GBDoesn’t fit your hardware
QLoRA≈ 117.3 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

This model isn’t on any leaderboard we cite yet.

  1. New language model

Sources

Some values on this page are still waiting to be checked against official pages: a filled dot is verified, a hollow one is not yet (with or without a source on file). How we verify data

Tencent · China · tencent.com

Snapshots, the same family and other models from this provider