Our latest quantization method introduces precision-adaptive quantization for ultra-low-bit models (1-2 bit), with benchmark-proven improvements on Llama-3-8B. This approach uses layer-specific strategies to preserve accuracy while maintaining extreme memory efficiency.

Benchmark Context

All tests conducted on Llama-3-8B-Instruct using:

Standard perplexity evaluation pipeline
2048-token context window
Same prompt set across all quantizations

Method

Dynamic Precision Allocation:
- First/Last 25% of layers → IQ4_XS (selected layers)
- Middle 50% → IQ2_XXS/IQ3_S (increase efficiency)
Critical Component Protection:
- Embeddings/output layers use Q5_K
- Reduces error propagation by 38% vs standard 1-2bit

Quantization Performance Comparison (Llama-3-8B)

Quantization	Standard PPL	DynamicGate PPL	Δ PPL	Std Size	DG Size	Δ Size	Std Speed	DG Speed
IQ2_XXS	11.30	9.84	-12.9%	2.5G	2.6G	+0.1G	234s	246s
IQ2_XS	11.72	11.63	-0.8%	2.7G	2.8G	+0.1G	242s	246s
IQ2_S	14.31	9.02	-36.9%	2.7G	2.9G	+0.2G	238s	244s
IQ1_M	27.46	15.41	-43.9%	2.2G	2.5G	+0.3G	206s	212s
IQ1_S	53.07	32.00	-39.7%	2.1G	2.4G	+0.3G	184s	209s

Key:

PPL = Perplexity (lower is better)
Δ PPL = Percentage change from standard to DynamicGate
Speed = Inference time (CPU avx2, 2048 token context)
Size differences reflect mixed quantization overhead

Key Improvements:

🔥 IQ1_M shows massive 43.9% perplexity reduction (27.46 → 15.41)
🚀 IQ2_S cuts perplexity by 36.9% while adding only 0.2GB
⚡ IQ1_S maintains 39.7% better accuracy despite 1-bit quantization

Tradeoffs:

All variants have modest size increases (0.1-0.3GB)
Inference speeds remain comparable (<5% difference)

When to Use These Models

📌 Fitting models into GPU VRAM

✔ Memory-constrained deployments

✔ Cpu and Edge Devices where 1-2bit errors can be tolerated

✔ Research into ultra-low-bit quantization

Choosing the Right Model Format

Selecting the correct model format depends on your hardware capabilities and memory constraints.

BF16 (Brain Float 16) – Use if BF16 acceleration is available

A 16-bit floating-point format designed for faster computation while retaining good precision.
Provides similar dynamic range as FP32 but with lower memory usage.
Recommended if your hardware supports BF16 acceleration (check your device's specs).
Ideal for high-performance inference with reduced memory footprint compared to FP32.

📌 Use BF16 if:
✔ Your hardware has native BF16 support (e.g., newer GPUs, TPUs).
✔ You want higher precision while saving memory.
✔ You plan to requantize the model into another format.

📌 Avoid BF16 if:
❌ Your hardware does not support BF16 (it may fall back to FP32 and run slower).
❌ You need compatibility with older devices that lack BF16 optimization.

F16 (Float 16) – More widely supported than BF16

A 16-bit floating-point high precision but with less of range of values than BF16.
Works on most devices with FP16 acceleration support (including many GPUs and some CPUs).
Slightly lower numerical precision than BF16 but generally sufficient for inference.

📌 Use F16 if:
✔ Your hardware supports FP16 but not BF16.
✔ You need a balance between speed, memory usage, and accuracy.
✔ You are running on a GPU or another device optimized for FP16 computations.

📌 Avoid F16 if:
❌ Your device lacks native FP16 support (it may run slower than expected).
❌ You have memory limitations.

Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference

Quantization reduces model size and memory usage while maintaining as much accuracy as possible.

Lower-bit models (Q4_K) → Best for minimal memory usage, may have lower precision.
Higher-bit models (Q6_K, Q8_0) → Better accuracy, requires more memory.

📌 Use Quantized Models if:
✔ You are running inference on a CPU and need an optimized model.
✔ Your device has low VRAM and cannot load full-precision models.
✔ You want to reduce memory footprint while keeping reasonable accuracy.

📌 Avoid Quantized Models if:
❌ You need maximum accuracy (full-precision models are better for this).
❌ Your hardware has enough VRAM for higher-precision formats (BF16/F16).

Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)

These models are optimized for extreme memory efficiency, making them ideal for low-power devices or large-scale deployments where memory is a critical constraint.

IQ3_XS: Ultra-low-bit quantization (3-bit) with extreme memory efficiency.
- Use case: Best for ultra-low-memory devices where even Q4_K is too large.
- Trade-off: Lower accuracy compared to higher-bit quantizations.
IQ3_S: Small block size for maximum memory efficiency.
- Use case: Best for low-memory devices where IQ3_XS is too aggressive.
IQ3_M: Medium block size for better accuracy than IQ3_S.
- Use case: Suitable for low-memory devices where IQ3_S is too limiting.
Q4_K: 4-bit quantization with block-wise optimization for better accuracy.
- Use case: Best for low-memory devices where Q6_K is too large.
Q4_0: Pure 4-bit quantization, optimized for ARM devices.
- Use case: Best for ARM-based devices or low-memory environments.

Summary Table: Model Format Selection

Model Format	Precision	Memory Usage	Device Requirements	Best Use Case
BF16	Highest	High	BF16-supported GPU/CPUs	High-speed inference with reduced memory
F16	High	High	FP16-supported devices	GPU inference when BF16 isn't available
Q4_K	Medium Low	Low	CPU or Low-VRAM devices	Best for memory-constrained environments
Q6_K	Medium	Moderate	CPU with more memory	Better accuracy while still being quantized
Q8_0	High	Moderate	CPU or GPU with enough VRAM	Best accuracy among quantized models
IQ3_XS	Very Low	Very Low	Ultra-low-memory devices	Extreme memory efficiency and low accuracy
Q4_0	Low	Low	ARM or low-memory devices	llama.cpp can optimize for ARM devices

Included Files & Details

`LiveCC-7B-Instruct-bf16.gguf`

Model weights preserved in BF16.
Use this if you want to requantize the model into a different format.
Best if your device supports BF16 acceleration.

`LiveCC-7B-Instruct-f16.gguf`

Model weights stored in F16.
Use if your device supports FP16, especially if BF16 is not available.

`LiveCC-7B-Instruct-bf16-q8_0.gguf`

Output & embeddings remain in BF16.
All other layers quantized to Q8_0.
Use if your device supports BF16 and you want a quantized version.

`LiveCC-7B-Instruct-f16-q8_0.gguf`

Output & embeddings remain in F16.
All other layers quantized to Q8_0.

`LiveCC-7B-Instruct-q4_k.gguf`

Output & embeddings quantized to Q8_0.
All other layers quantized to Q4_K.
Good for CPU inference with limited memory.

`LiveCC-7B-Instruct-q4_k_s.gguf`

Smallest Q4_K variant, using less memory at the cost of accuracy.
Best for very low-memory setups.

`LiveCC-7B-Instruct-q6_k.gguf`

Output & embeddings quantized to Q8_0.
All other layers quantized to Q6_K .

`LiveCC-7B-Instruct-q8_0.gguf`

Fully Q8 quantized model for better accuracy.
Requires more memory but offers higher precision.

`LiveCC-7B-Instruct-iq3_xs.gguf`

IQ3_XS quantization, optimized for extreme memory efficiency.
Best for ultra-low-memory devices.

`LiveCC-7B-Instruct-iq3_m.gguf`

IQ3_M quantization, offering a medium block size for better accuracy.
Suitable for low-memory devices.

`LiveCC-7B-Instruct-q4_0.gguf`

Pure Q4_0 quantization, optimized for ARM devices.
Best for low-memory environments.
Prefer IQ4_NL for better accuracy.

🚀 If you find these models useful

❤ Please click "Like" if you find this useful!
Help me test my AI-Powered Network Monitor Assistant with quantum-ready security checks:
👉 Quantum Network Monitor

💬 How to test:

Click the chat icon (bottom right on any page)
Choose an AI assistant type:
- TurboLLM (GPT-4-mini)
- FreeLLM (Open-source)
- TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

Function calling against live network services
How small can a model go while still handling:
- Automated Nmap scans
- Quantum-readiness checks
- Metasploit integration

🟡 TestLLM – Current experimental model (llama.cpp on 6 CPU threads):

✅ Zero-configuration setup
⏳ 30s load time (slow inference but no API costs)
🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4-mini for:

Real-time network diagnostics
Automated penetration testing (Nmap/Metasploit)
🔑 Get more tokens by downloading our Quantum Network Monitor Agent

🔵 HugLLM – Open-source models (≈8B params):

2x more tokens than TurboLLM
AI-powered log analysis
🌐 Runs on Hugging Face Inference API

💡 Example AI Commands to Test:

"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a quick Nmap vulnerability test"
'"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code from. This is a very flexible and powerful feature. Use with caution!

Final word

I fund the servers to create the models files, run the Quantum Network Monitor Service and Pay for Inference from Novita and OpenAI all from my own pocket. All of the code for creating the models and the work I have done with Quantum Network Monitor is open source. Feel free to use what you find useful. Please support my work and consider buying me a coffee . This will help me pay for the services and increase the token limits for everyone.

Thank you :)

LiveCC-7B-Instruct

Introduction

We introduce LiveCC, the first video LLM capable of real-time commentary, trained with a novel video-ASR streaming method, SOTA on both streaming and offline benchmarks.

Project Page: https://showlab.github.io/livecc

[!Important] This is the SFT model. The base model is at LiveCC-7B-Base.

Training with Streaming Frame-Words Paradigm

Quickstart

Gradio Demo

Please refer to https://github.com/showlab/livecc:

Hands-on

Like qwen-vl-utils, we offer a toolkit to help you handle various types of visual input more conveniently, especially on video streaming inputs. You can install it using the following command:

pip install qwen-vl-utils livecc-utils liger_kernel

Here we show a code snippet to show you how to do real-time video commentary with transformers and the above utils:

python

import functools, torch, os, tqdm
from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
apply_liger_kernel_to_qwen2_vl() # important. our model is trained with this. keep consistency
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, LogitsProcessor, logging
from livecc_utils import prepare_multiturn_multimodal_inputs_for_generation, get_smart_resized_clip, get_smart_resized_video_reader
from qwen_vl_utils import process_vision_info

class LiveCCDemoInfer:
  fps = 2
  initial_fps_frames = 6
  streaming_fps_frames = 2
  initial_time_interval = initial_fps_frames / fps
  streaming_time_interval = streaming_fps_frames / fps
  frame_time_interval = 1 / fps
  def __init__(self, model_path: str = None, device_id: int = 0):
      self.model = Qwen2VLForConditionalGeneration.from_pretrained(
          model_path, torch_dtype="auto", 
          device_map=f'cuda:{device_id}', 
          attn_implementation='flash_attention_2'
      )
      self.processor = AutoProcessor.from_pretrained(model_path, use_fast=False)
      self.model.prepare_inputs_for_generation = functools.partial(prepare_multiturn_multimodal_inputs_for_generation, self.model)
      message = {
          "role": "user",
          "content": [
              {"type": "text", "text": 'livecc'},
          ]
      }
      texts = self.processor.apply_chat_template([message], tokenize=False)
      self.system_prompt_offset = texts.index('<|im_start|>user')
      self._cached_video_readers_with_hw = {}


  def live_cc(
      self,
      query: str,
      state: dict,
      max_pixels: int = 384 * 28 * 28,
      default_query: str = 'Please describe the video.',
      do_sample: bool = True,
      repetition_penalty: float = 1.05,
      **kwargs,
  ): 
      """
      state: dict, (maybe) with keys:
          video_path: str, video path
          video_timestamp: float, current video timestamp
          last_timestamp: float, last processed video timestamp
          last_video_pts_index: int, last processed video frame index
          video_pts: np.ndarray, video pts
          last_history: list, last processed history
          past_key_values: llm past_key_values
          past_ids: past generated ids
      """
      # 1. preparation: video_reader, and last processing info
      video_timestamp, last_timestamp = state.get('video_timestamp', 0), state.get('last_timestamp', -1 / self.fps)
      video_path = state['video_path']
      if video_path not in self._cached_video_readers_with_hw:
          self._cached_video_readers_with_hw[video_path] = get_smart_resized_video_reader(video_path, max_pixels)
          video_reader = self._cached_video_readers_with_hw[video_path][0]
          video_reader.get_frame_timestamp(0)
          state['video_pts'] = torch.from_numpy(video_reader._frame_pts[:, 1])
          state['last_video_pts_index'] = -1
      video_pts = state['video_pts']
      if last_timestamp + self.frame_time_interval > video_pts[-1]:
          state['video_end'] = True
          return 
      video_reader, resized_height, resized_width = self._cached_video_readers_with_hw[video_path]
      last_video_pts_index = state['last_video_pts_index']

      # 2. which frames will be processed
      initialized = last_timestamp >= 0
      if not initialized:
          video_timestamp = max(video_timestamp, self.initial_time_interval)
      if video_timestamp <= last_timestamp + self.frame_time_interval:
          return
      timestamps = torch.arange(last_timestamp + self.frame_time_interval, video_timestamp, self.frame_time_interval) # add compensation
      
      # 3. fetch frames in required timestamps
      clip, clip_timestamps, clip_idxs = get_smart_resized_clip(video_reader, resized_height, resized_width, timestamps, video_pts, video_pts_index_from=last_video_pts_index+1)
      state['last_video_pts_index'] = clip_idxs[-1]
      state['last_timestamp'] = clip_timestamps[-1]

      # 4. organize to interleave frames
      interleave_clips, interleave_timestamps = [], []
      if not initialized:
          interleave_clips.append(clip[:self.initial_fps_frames])
          interleave_timestamps.append(clip_timestamps[:self.initial_fps_frames])
          clip = clip[self.initial_fps_frames:]
          clip_timestamps = clip_timestamps[self.initial_fps_frames:]
      if len(clip) > 0:
          interleave_clips.extend(list(clip.split(self.streaming_fps_frames)))
          interleave_timestamps.extend(list(clip_timestamps.split(self.streaming_fps_frames)))

      # 5. make conversation and send to model
      for clip, timestamps in zip(interleave_clips, interleave_timestamps):
          start_timestamp, stop_timestamp = timestamps[0].item(), timestamps[-1].item() + self.frame_time_interval
          message = {
              "role": "user",
              "content": [
                  {"type": "text", "text": f'Time={start_timestamp:.1f}-{stop_timestamp:.1f}s'},
                  {"type": "video", "video": clip}
              ]
          }
          if not query and not state.get('query', None):
              query = default_query
              print(f'No query provided, use default_query={default_query}')
          if query and state.get('query', None) != query:
              message['content'].append({"type": "text", "text": query})
              state['query'] = query
          texts = self.processor.apply_chat_template([message], tokenize=False, add_generation_prompt=True, return_tensors='pt')
          past_ids = state.get('past_ids', None)
          if past_ids is not None:
              texts = '<|im_end|>\n' + texts[self.system_prompt_offset:]
          inputs = self.processor(
              text=texts,
              images=None,
              videos=[clip],
              return_tensors="pt",
              return_attention_mask=False
          )
          inputs.to('cuda')
          if past_ids is not None:
              inputs['input_ids'] = torch.cat([past_ids, inputs.input_ids], dim=1) 
          outputs = self.model.generate(
              **inputs, past_key_values=state.get('past_key_values', None), 
              return_dict_in_generate=True, do_sample=do_sample, 
              repetition_penalty=repetition_penalty,
          )
          state['past_key_values'] = outputs.past_key_values
          state['past_ids'] = outputs.sequences[:, :-1]
          yield (start_timestamp, stop_timestamp), self.processor.decode(outputs.sequences[0, inputs.input_ids.size(1):], skip_special_tokens=True), state

model_path = 'chenjoya/LiveCC-7B-Instruct'
# download a test video at: https://github.com/showlab/livecc/blob/main/demo/sources/howto_fix_laptop_mute_1080p.mp4
video_path = "demo/sources/howto_fix_laptop_mute_1080p.mp4"
query = "Please describe the video."

infer = LiveCCDemoInfer(model_path=model_path)
state = {'video_path': video_path}
commentaries = []
t = 0
for t in range(31):
    state['video_timestamp'] = t
    for (start_t, stop_t), response, state in infer.live_cc(
        query=query, state=state, 
        max_pixels = 384 * 28 * 28, repetition_penalty=1.05, 
        streaming_eos_base_threshold=0.0, streaming_eos_threshold_step=0
    ):
        print(f'{start_t}s-{stop_t}s: {response}')
        commentaries.append([start_t, stop_t, response])
    if state.get('video_end', False):
        break
    t += 1

Here we show a code snippet to show you how to do common video (multi-turn) qa with transformers and the above utils:

python

import functools, torch
from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
apply_liger_kernel_to_qwen2_vl() # important. our model is trained with this. keep consistency
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, LogitsProcessor, logging
from livecc_utils import prepare_multiturn_multimodal_inputs_for_generation, get_smart_resized_clip, get_smart_resized_video_reader
from qwen_vl_utils import process_vision_info

class LiveCCDemoInfer:
  fps = 2
  initial_fps_frames = 6
  streaming_fps_frames = 2
  initial_time_interval = initial_fps_frames / fps
  streaming_time_interval = streaming_fps_frames / fps
  frame_time_interval = 1 / fps

  def __init__(self, model_path: str = None, device: str = 'cuda'):
      self.model = Qwen2VLForConditionalGeneration.from_pretrained(
          model_path, torch_dtype="auto", 
          device_map=device, 
          attn_implementation='flash_attention_2'
      )
      self.processor = AutoProcessor.from_pretrained(model_path, use_fast=False)
      self.streaming_eos_token_id = self.processor.tokenizer(' ...').input_ids[-1]
      self.model.prepare_inputs_for_generation = functools.partial(prepare_multiturn_multimodal_inputs_for_generation, self.model)
      message = {
          "role": "user",
          "content": [
              {"type": "text", "text": 'livecc'},
          ]
      }
      texts = self.processor.apply_chat_template([message], tokenize=False)
      self.system_prompt_offset = texts.index('<|im_start|>user')

  def video_qa(
      self,
      message: str,
      state: dict,
      do_sample: bool = True,
      repetition_penalty: float = 1.05,
      **kwargs,
  ): 
      """
      state: dict, (maybe) with keys:
          video_path: str, video path
          video_timestamp: float, current video timestamp
          last_timestamp: float, last processed video timestamp
          last_video_pts_index: int, last processed video frame index
          video_pts: np.ndarray, video pts
          last_history: list, last processed history
          past_key_values: llm past_key_values
          past_ids: past generated ids
      """
      video_path = state.get('video_path', None)
      conversation = []
      past_ids = state.get('past_ids', None)
      content = [{"type": "text", "text": message}]
      if past_ids is None and video_path: # only use once
          content.insert(0, {"type": "video", "video": video_path})
      conversation.append({"role": "user", "content": content})
      image_inputs, video_inputs = process_vision_info(conversation)
      texts = self.processor.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True, return_tensors='pt')
      if past_ids is not None:
          texts = '<|im_end|>\n' + texts[self.system_prompt_offset:]
      inputs = self.processor(
          text=texts,
          images=image_inputs,
          videos=video_inputs,
          return_tensors="pt",
          return_attention_mask=False
      )
      inputs.to(self.model.device)
      if past_ids is not None:
          inputs['input_ids'] = torch.cat([past_ids, inputs.input_ids], dim=1) 
      outputs = self.model.generate(
          **inputs, past_key_values=state.get('past_key_values', None), 
          return_dict_in_generate=True, do_sample=do_sample, 
          repetition_penalty=repetition_penalty,
          max_new_tokens=512,
      )
      state['past_key_values'] = outputs.past_key_values
      state['past_ids'] = outputs.sequences[:, :-1]
      response = self.processor.decode(outputs.sequences[0, inputs.input_ids.size(1):], skip_special_tokens=True)
      return response, state

model_path = 'chenjoya/LiveCC-7B-Instruct'
# download a test video at: https://github.com/showlab/livecc/blob/main/demo/sources/howto_fix_laptop_mute_1080p.mp4
video_path = "demo/sources/howto_fix_laptop_mute_1080p.mp4"

infer = LiveCCDemoInfer(model_path=model_path)
state = {'video_path': video_path}
# first round
query1 = 'What is the video?'
response1, state = infer.video_qa(message=query1, state=state)
print(f'Q1: {query1}\nA1: {response1}')
# second round
query2 = 'How do you know that?'
response2, state = infer.video_qa(message=query2, state=state)
print(f'Q2: {query2}\nA2: {response2}')

Performance

Limitations

This model is finetuned on LiveCC-7B-Base, which is starting from Qwen2-VL-7B-Base, so it may have limitations mentioned in https://huggingface.co/Qwen/Qwen2-VL-7B.
When performing real-time video commentary, it may appear collapse --- e.g., repeat pattern. If you encounter this situation, try to adjust repetition_penalty, streaming_eos_base_threshold, and streaming_eos_threshold_step.
This model only has a context window of 32768. Using more visual tokens per frame (e.g. 768 * 28 * 28) will have better performance, but will shorten the working duration.

These limitations serve as ongoing directions for model optimization and improvement, and we are committed to continually enhancing the model's performance and scope of application.

Citation

If you find our work helpful, feel free to give us a cite.

@article{livecc,
  author       = {Joya Chen and Ziyun Zeng and Yiqi Lin and Wei Li and Zejun Ma and Mike Zheng Shou},
  title        = {LiveCC: Learning Video LLM with Streaming Speech Transcription at Scale},
  journal      = {arXiv preprint arXiv:2504.16030}
  year         = {2025},
}

Mungert/LiveCC-7B-Instruct-GGUF

LiveCC-7B-Instruct GGUF Models

Model Generation Details

Ultra-Low-Bit Quantization with IQ-DynamicGate (1-2 bit)