airllm

Quickstart | Configurations | MacOS | Example notebooks | FAQ

AirLLM dramatically reduces inference memory usage, letting 70B large language models run on a single 4GB GPU card — without quantization, distillation, or pruning. You can even run Kimi K3 (2.8T) — the largest open-source model released to date — on under 4GB, Qwen3.8-Flash-Next (125B) on 6GB, and DeepSeek-V3 (671B) on ~12GB. We now also support training huge models on small VRAM: Qwen3.8-Flash-Next (125B) under 6GB.

AI Agents Recommendation:

• Best AI Game Sprite Generator

• Best AI Facial Expression Editor

Updates

[2026/09] Training support: stream frozen weights one layer at a time and keep adapters on the GPU. Qwen3.8-Flash-Next (125B) trains under 6GB (RTX 3060 Ti); Qwen3.8-27B trains in ~2GB at seq 512. See Training.

[2026/08] Qwen3.8-Flash-Next support: Qwen's 125B MoE flagship (Qwen4ExpForConditionalGeneration) with a ~51B n-gram embedding runs in 5.95GB of VRAM, measured end to end on one RTX 4090. The n-gram table is file-mapped on the host (a 64GB machine is enough); decoder layers stream. Needs a transformers build with in-tree qwen4_exp (pip install git+https://github.com/huggingface/transformers.git today) and ~360GB of checkpoint disk (delete_original=True reclaims the originals after the split).

[2026/08] Qwen3.8-27B support: Qwen's new dense VL (Gated DeltaNet + Gated Attention, native vision) runs in 3.33GB of VRAM, measured end to end on one RTX 3090. Needs transformers 5.8+.

[2026/07] Kimi K3 (2.8T) support: the largest open-source model runs on a single card in 3.72GB of VRAM, measured end to end on one RTX 6000 Ada. Per-expert streaming loads only the experts a token actually routes to. K3 brings three requirements of its own: pip install compressed-tensors flash-attn (its model code mandates flash attention regardless of what you request), a CUDA 12 build of torch, since no prebuilt flash-attn wheel exists for CUDA 13 yet, and transformers 4.56.x, as its remote code does not load on 5.x.

[2026/06] v3.0: FP8 model support + the latest models. Run DeepSeek-V3 (671B) on ~12GB and Qwen3-235B on ~3GB, plus Qwen3, Llama 3.x/4, DeepSeek V2/V3, Phi-4, Gemma and more — all through a single AutoModel.

[2024/08/20] v2.11.0: Support Qwen2.5

[2024/08/18] v2.10.1 Support CPU inference. Support non sharded models. Thanks @NavodPeiris for the great work!

[2024/07/30] Support Llama3.1 405B (example notebook). Support 8bit/4bit quantization.

[2024/04/20] AirLLM supports Llama3 natively already. Run Llama3 70B on 4GB single GPU.

[2023/12/25] v2.8.2: Support MacOS running 70B large language models.

[2023/12/20] v2.7: Support AirLLMMixtral.

[2023/12/20] v2.6: Added AutoModel, automatically detect model type, no need to provide model class to initialize model.

[2023/12/18] v2.5: added prefetching to overlap the model loading and compute. 10% speed improvement.

[2023/12/03] added support of ChatGLM, QWen, Baichuan, Mistral, InternLM!

[2023/12/02] added support for safetensors. Now support all top 10 models in open llm leaderboard.

[2023/12/01] airllm 2.0. Support compressions: 3x run time speed up!

[2023/11/20] airllm Initial version!

Star History

Table of Contents

• Quick start
• Model Compression
• Configurations
• Run on MacOS
• Example notebooks
• Supported Models
• Training
• Acknowledgement
• FAQ

Quickstart

1. Install package

First, install the airllm pip package.

pip install airllm

2. Inference

Then, initialize AirLLMLlama2, pass in the huggingface repo ID of the model being used, or the local path, and inference can be performed similar to a regular transformer model.

(You can also specify the path to save the split layered model through layer_shards_saving_path when init AirLLMLlama2.

from airllm import AutoModel

MAX_LENGTH = 128
# just pass a hugging face repo id — works with almost any popular model:
model = AutoModel.from_pretrained("Qwen/Qwen3-32B")

# go bigger with the exact same one line:
#model = AutoModel.from_pretrained("Qwen/Qwen3.8-27B")          # 27B dense VL, 3.33GB
#model = AutoModel.from_pretrained("Qwen/Qwen3.8-Flash-Next")    # 125B MoE + 51B PLE, 5.95GB
#model = AutoModel.from_pretrained("Qwen/Qwen3-235B-A22B")     # 235B, runs in ~3GB
#model = AutoModel.from_pretrained("deepseek-ai/DeepSeek-V3")  # 671B, runs in ~12GB

# or use a model's local path...
#model = AutoModel.from_pretrained("/home/ubuntu/.cache/huggingface/hub/models--Qwen--Qwen3-32B/snapshots/...")

input_text = [
        'What is the capital of United States?',
        #'I like',
    ]

input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH, 
    padding=False)
           
generation_output = model.generate(
    input_tokens['input_ids'].to(model.device), 
    max_new_tokens=20,
    use_cache=True,
    return_dict_in_generate=True)

output = model.tokenizer.decode(generation_output.sequences[0])

print(output)

Note: During inference, the original model will first be decomposed and saved layer-wise. Please ensure there is sufficient disk space in the huggingface cache directory.

Model Compression - 3x Inference Speed Up!

We just added model compression based on block-wise quantization-based model compression. Which can further speed up the inference speed for up to 3x , with almost ignorable accuracy loss! (see more performance evaluation and why we use block-wise quantization in this paper)

How to enable model compression speed up:

• Step 1. make sure you have bitsandbytes installed by pip install -U bitsandbytes
• Step 2. make sure airllm version later than 2.0.0: pip install -U airllm
• Step 3. when initialize the model, passing the argument compression ('4bit' or '8bit'):

model = AutoModel.from_pretrained("garage-bAInd/Platypus2-70B-instruct",
                     compression='4bit' # specify '8bit' for 8-bit block-wise quantization 
                    )

What are the differences between model compression and quantization?

Quantization normally needs to quantize both weights and activations to really speed things up. Which makes it harder to maintain accuracy and avoid the impact of outliers in all kinds of inputs.

While in our case the bottleneck is mainly at the disk loading, we only need to make the model loading size smaller. So, we get to only quantize the weights' part, which is easier to ensure the accuracy.

Configurations

When initialize the model, we support the following configurations:

• compression: supported options: 4bit, 8bit for 4-bit or 8-bit block-wise quantization, or by default None for no compression
• profiling_mode: supported options: True to output time consumptions or by default False
• layer_shards_saving_path: optionally another path to save the split model
• hf_token: huggingface token can be provided here if downloading gated models like: meta-llama/Llama-2-7b-hf
• prefetching: prefetching to overlap the model loading and compute. By default, turned on. For now, only AirLLMLlama2 supports this.
• delete_original: if you don't have too much disk space, you can set delete_original to true to delete the original downloaded hugging face model, only keep the transformed one to save half of the disk space.

MacOS

Just install airllm and run the code the same as on linux. See more in Quick Start.

• make sure you installed mlx and torch
• you probably need to install python native see more here
• only Apple silicon is supported
• only Llama-style models (LlamaForCausalLM, MistralForCausalLM) run on macOS for now; other architectures raise an error before anything is downloaded

Example [python notebook] (https://github.com/lyogavin/airllm/blob/main/air_llm/examples/run_on_macos.ipynb)

Example Python Notebook

Example colabs here:

example of other models (ChatGLM, QWen, Baichuan, Mistral, etc):

• ChatGLM:

from airllm import AutoModel
MAX_LENGTH = 128
model = AutoModel.from_pretrained("THUDM/chatglm3-6b-base")
input_text = ['What is the capital of China?',]
input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH, 
    padding=True)
generation_output = model.generate(
    input_tokens['input_ids'].to(model.device), 
    max_new_tokens=5,
    use_cache= True,
    return_dict_in_generate=True)
model.tokenizer.decode(generation_output.sequences[0])

• QWen:

from airllm import AutoModel
MAX_LENGTH = 128
model = AutoModel.from_pretrained("Qwen/Qwen-7B")
input_text = ['What is the capital of China?',]
input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH)
generation_output = model.generate(
    input_tokens['input_ids'].to(model.device), 
    max_new_tokens=5,
    use_cache=True,
    return_dict_in_generate=True)
model.tokenizer.decode(generation_output.sequences[0])

• Baichuan, InternLM, Mistral, etc:

from airllm import AutoModel
MAX_LENGTH = 128
model = AutoModel.from_pretrained("baichuan-inc/Baichuan2-7B-Base")
#model = AutoModel.from_pretrained("internlm/internlm-20b")
#model = AutoModel.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
input_text = ['What is the capital of China?',]
input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH)
generation_output = model.generate(
    input_tokens['input_ids'].to(model.device), 
    max_new_tokens=5,
    use_cache=True,
    return_dict_in_generate=True)
model.tokenizer.decode(generation_output.sequences[0])

To request other model support: here

Supported Models

AirLLM works out of the box with virtually every popular open LLM — just pass its Hugging Face ID to AutoModel.from_pretrained(...). That covers all the major families:

Llama (2 / 3 / 3.1 / 3.3 / 4) · Qwen (1 / 2 / 2.5 / 3 / 3.5 / 3.8, including MoE, Flash-Next, FP8, and native VL) · DeepSeek (V2 / V3 / R1) · Mistral & Mixtral · Phi · Gemma · ChatGLM · Baichuan · InternLM · Yi · Kimi K3 — and most new models the day they're released.

Tiny GPU, huge models

The trick: AirLLM only ever keeps one layer on the GPU at a time, so the VRAM you need depends on the model's layer size — not its total size. That's how a 671B model fits on a hobbyist card:

Model	Size	GPU VRAM
Qwen3 / Mistral / Phi (≈8B)	8B	~1–2 GB
Qwen3-30B / Mixtral (MoE)	30–47B	~1–3 GB
Qwen3.8-27B (dense VL)	27B	3.33 GB
Qwen3.8-Flash-Next (MoE + PLE)	~180B	5.95 GB
Qwen3-235B (MoE)	235B	~3 GB
Llama 3.x 70B (full precision)	70B	~4 GB
Llama 3.1 405B	405B	~8 GB
DeepSeek-V3	671B	~12 GB

Same one line of code for all of them — no special setup.

Training

AirLLM can fine-tune huge models on a small GPU. Frozen base weights stream from disk one decoder layer at a time; only the adapters stay resident. Qwen3.8-Flash-Next (125B) trains under 6GB; Qwen3.8-27B trains in ~2GB at seq 512.

This is not Hugging Face Trainer / bitsandbytes QLoRA. Flash-Next needs a transformers build with in-tree qwen4_exp (pip install git+https://github.com/huggingface/transformers.git today).

1. Prepare a dataset

One JSON object per line (.jsonl). The usual field is text — next-token prediction over the whole string:

{"text": "Your first training document. Can be a few sentences or a few paragraphs."}
{"text": "Your second training document."}

Instruction pairs work too. Loss is applied on the completion only:

{"prompt": "What is AirLLM?", "completion": "A library that runs and trains huge models on small VRAM."}
{"instruction": "Translate to English", "input": "bonjour", "output": "hello"}

A .txt file is also fine: one example per blank-line-separated block. A two-line starter file lives at air_llm/examples/sft_example.jsonl.

2. Run training

From the repo root, point --data at your file:

python air_llm/examples/train_qwen38_flash_next_lora.py \
  --data my_data.jsonl \
  --seq-len 512 \
  --epochs 1 \
  --save-adapter qwen38-flash-next-lora.pt

For the 27B dense model:

python air_llm/examples/train_qwen38_lora.py \
  --data my_data.jsonl \
  --seq-len 512 \
  --epochs 1 \
  --save-adapter qwen38-27b-lora.pt

--steps N stops after N examples (useful for a smoke test). Omit --data and the script overfits a built-in snippet.

Python API

from airllm import AirLLMLoRAQwen4Exp

trainer = AirLLMLoRAQwen4Exp(
    "Qwen/Qwen3.8-Flash-Next",
    max_seq_len=512,
    lora_r=16,
    delete_original=True,
)

tok = trainer.tokenizer
if tok.pad_token_id is None:
    tok.pad_token = tok.eos_token

encoded = tok(
    "Your training text here.",
    return_tensors="pt",
    truncation=True,
    max_length=512,
)
loss = trainer.train_step(
    encoded["input_ids"].to(model.device),
    attention_mask=encoded.get("attention_mask"),
)
print(loss)
trainer.save_adapter("qwen38-flash-next-lora.pt")

AirLLMLoRA is the same API for Qwen/Qwen3.8-27B.

Acknowledgement

A lot of the code are based on SimJeg's great work in the Kaggle exam competition. Big shoutout to SimJeg:

GitHub account @SimJeg, the code on Kaggle, the associated discussion.

FAQ

1. MetadataIncompleteBuffer

safetensors_rust.SafetensorError: Error while deserializing header: MetadataIncompleteBuffer

If you run into this error, most possible cause is you run out of disk space. The process of splitting model is very disk-consuming. See this. You may need to extend your disk space, clear huggingface .cache and rerun.

2. ValueError: max() arg is an empty sequence

Most likely you are loading QWen or ChatGLM model with Llama2 class. Try the following:

For QWen model:

from airllm import AutoModel #<----- instead of AirLLMLlama2
AutoModel.from_pretrained(...)

For ChatGLM model:

from airllm import AutoModel #<----- instead of AirLLMLlama2
AutoModel.from_pretrained(...)

3. 401 Client Error....Repo model ... is gated.

Some models are gated models, needs huggingface api token. You can provide hf_token:

model = AutoModel.from_pretrained("meta-llama/Llama-2-7b-hf", #hf_token='HF_API_TOKEN')

4. ValueError: Asking to pad but the tokenizer does not have a padding token.

Some model's tokenizer doesn't have padding token, so you can set a padding token or simply turn the padding config off:

input_tokens = model.tokenizer(input_text,
   return_tensors="pt", 
   return_attention_mask=False, 
   truncation=True, 
   max_length=MAX_LENGTH, 
   padding=False  #<-----------   turn off padding 
)

Citing AirLLM

If you find AirLLM useful in your research and wish to cite it, please use the following BibTex entry:

@software{airllm2023,
  author = {Gavin Li},
  title = {AirLLM: scaling large language models on low-end commodity computers},
  url = {https://github.com/lyogavin/airllm/},
  version = {0.0},
  year = {2023},
}

Contribution

Welcomed contributions, ideas and discussions!

If you find it useful, please ⭐ or buy me a coffee! 🙏