Skip to content
Video

Llama-V-JEPA-LM-Qwen3-30B-A3B

a powerful video-language model for scene dynamics analysis

2 min readupdated

The model combines a frozen V-JEPA 2.1 (ViT-G) video encoder with a Qwen3-30B-A3B MoE language model. It is designed to understand motion, direction, speed, and object states in video clips. Thanks to a dedicated spectral bridge, each clip is converted into 1,024 visual tokens regardless of the original resolution. The new model outperforms the base Qwen3-VL-30B-A3B and other open alternatives across a range of video benchmarks while retaining the strong mathematical and text skills of the original language backbone.

Key features

  • Outperforms open models of a similar class in video comprehension tests (69.49 average across 11 benchmarks).
  • Trained using counterfactual supervision for precise tracking of scene dynamics.
  • Efficient resolution handling: the number of visual tokens is fixed at 1,024 and does not depend on the original video resolution.
  • Retains the strong mathematical and text capabilities of the base language model (GSM8K 88.7).

What you can do

  • Analyze video fragments of any length using the built-in chat assistant to answer questions about the content.
  • Set up multiple-choice questions for accurate evaluation of logic and events in the video.
  • Run analysis for a specific video timecode using the start_sec and end_sec parameters.
  • Perform autonomous video encoding using the encode_video function and ask different questions of the same cache.

Example

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "kkt94/Llama-V-JEPA-LM-Qwen3-30B-A3B",
    trust_remote_code=True, dtype=torch.bfloat16,
).to("cuda").eval()

print(model.chat("my_video.mp4", "Describe what happens in the video."))
Read moreDetails
Architecture: MoE (30.53B parameters, 8 active experts out of 128) + V-JEPA 2.1 encoder (1.85B).
Hardware requirements: about 61 GB VRAM to process a single test clip.
License: Apache-2.0 (encoder: MIT, training data: Llama 3.2 Community License).
Discuss Telegram 13 51

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 60 GB · ready files72 GB
  • FP8 / INT8recommendedweights 30 GB38 GB
  • 4-bitweights 17 GB23 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # kkt94/Llama-V-JEPA-LM-Qwen3-30B-A3B
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download kkt94/Llama-V-JEPA-LM-Qwen3-30B-A3B --local-dir ~/models/Llama-V-JEPA-LM-Qwen3-30B-A3B
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (3)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • LightX2V step distill0.6.1from November 9, 2026new release

    5–10× fewer steps

    Distilled LoRAs and an engine: video in 4–8 steps instead of 30–50.

    How to enable: Load their step-distill LoRA, CFG 1.0, 4–8 steps.

  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. Qwen3-30B-A3B-Basefamily root
  2. Qwen3-30B-A3B
  3. Llama-V-JEPA-LM-Qwen3-30B-A3B

New posts about versions of Qwen3-30B-A3B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
Merges 7All on Hugging Face

About the model

kkt94/Llama-V-JEPA-LM-Qwen3-30B-A3B
Author
kkt94
Category
Video
Released
Posted

On NeyroLearn

Hugging Face

5113

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments