WeMM Video 4B NVFP4

Language Tencent Partly verified
  • Open weights

Prices

0 offerings · prices as listed

No priced offering is listed for this model yet.

Details

Family
wemm-video-nvfp

Run locally

Weights: tencent/WeMM-Video-4B-NVFP4
  • Parameters 5.6B
  • Dense
  • Native precision BF16
  • Layers 32
  • KV cache per token 32 KB
  • Config positions 256K (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
Runs well BF16 (native) (automatic) 8K context 61–92 tok/sOne GPU
WeightsKV cacheOverheadUsable
Needs
≈ 8.2 GB (weights 7.4 (maker’s file) + KV cache 0.3 + overhead 0.5)
Usable
21.6 GB (90% of its memory)
Generation speed
61–92 tok/s
Longest context
256K (the model’s limit)
First token
about 0.18–0.27 s for a 2K prompt
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file7.4 GB≈ 8.2Runs well 61–92 tok/s256K
Q8_05.6 GB≈ 6.3Runs well 81–121 tok/s256K
FP85.2 GB≈ 6.0Runs well 86–129 tok/s256K
Q6_K4.3 GB≈ 5.0Runs well 103–155 tok/s256K
Q5_K_M3.7 GB≈ 4.5Runs well 118–177 tok/s256K
Q5_03.6 GB≈ 4.3Runs well 122–183 tok/s256K
Q4_K_M3.2 GB≈ 3.9Runs well 137–206 tok/s256K
Q4_02.9 GB≈ 3.7Runs well 147–221 tok/s256K
IQ4_XS2.8 GB≈ 3.5Runs well 155–233 tok/s256K
Q3_K_Mquality drops2.6 GB≈ 3.3Runs well 167–251 tok/s256K
IQ3_XXSquality drops2.0 GB≈ 2.7Runs well 209–313 tok/s256K
Q2_Kquality drops2.0 GB≈ 2.7Runs well 212–319 tok/s256K
IQ2_XXSquality drops1.3 GB≈ 2.1Runs well 294–441 tok/s256K

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 GeForce RTX 3080; unified memory Apple M5 · 16 GB.

Run commands

vLLM
vllm serve tencent/WeMM-Video-4B-NVFP4 --max-model-len 8192
Memory to fine-tune
Full≈ 86.5 GBDoesn’t fit your hardware
LoRA≈ 13.5 GBFits your hardware
QLoRA≈ 6.0 GBFits 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