selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

empero-ai logo

Can the RTX 4070 run Qwen3.8-4B-Distill?

Qwen3.8-4B-Distill·4.7B·GeForce RTX 4070·12 GB VRAM

Saved in this browser, used on every model page.
Yes

Yes. The RTX 4070 runs Qwen3.8-4B-Distill fully in VRAM at FP16 / BF16 (about 10.0 GB of 12 GB), at roughly 30 tokens/s.

best quant
FP16 / BF16
needs
10.0 GB
usable
12 GB
est. speed
~30 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the Qwen3.8-4B-Distill page →

Every quantisation of Qwen3.8-4B-Distill on the RTX 4070

QuantQualityMemoryFitSpeed
Q2_K Heavy loss3.4 GBfits~181 tok/s
Q3_K_M Noticeable loss3.8 GBfits~122 tok/s
Q4_K_M Small loss4.2 GBfits~98 tok/s
Q5_K_M Near-lossless4.6 GBfits~84 tok/s
Q6_K Near-lossless5.0 GBfits~72 tok/s
Q8_0 Lossless6.0 GBfits~56 tok/s
FP16 / BF16 pickLossless10.0 GBfits~30 tok/s

Speed is a bandwidth-based estimate for generating text (504 GB/s); prompt processing and long contexts are slower.

Measured speed on the RTX 4070

No community reports for this pair yet.

Report your own speed →

Run Qwen3.8-4B-Distill on the RTX 4070

Using FP16 / BF16, the best fit for RTX 4070.

Ollama · BF16 straight from Hugging Facedocs
ollama run hf.co/empero-ai/Qwen3.8-4B-Distill-GGUF:BF16
llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf empero-ai/Qwen3.8-4B-Distill-GGUF:BF16

Downloads the GGUF on first run. Installed through llama.app? The same flags work with `llama serve`.

LM Studio · Download with the lms CLIdocs
lms get empero-ai/Qwen3.8-4B-Distill-GGUF@BF16

Then load it from the LM Studio app or with `lms load`.

vLLM · Serve an OpenAI-compatible APIdocs
vllm serve empero-ai/Qwen3.8-4B-Distill
vLLM · Docker Compose (NVIDIA)docs
services:
  vllm:
    image: vllm/vllm-openai:latest
    command: empero-ai/Qwen3.8-4B-Distill
    ports:
      - "8000:8000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 8000.

SGLang · Serve an OpenAI-compatible APIdocs
sglang serve empero-ai/Qwen3.8-4B-Distill --host 0.0.0.0 --port 30000
SGLang · Docker Compose (NVIDIA)docs
services:
  sglang:
    image: lmsysorg/sglang:latest
    command: sglang serve empero-ai/Qwen3.8-4B-Distill --host 0.0.0.0 --port 30000
    ports:
      - "30000:30000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 30000.

Transformers · Chat with the pipeline APIdocs
from transformers import pipeline

pipe = pipeline("text-generation", model="empero-ai/Qwen3.8-4B-Distill", dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Explain self-hosting in one paragraph."}]
out = pipe(messages, max_new_tokens=512)
print(out[0]["generated_text"][-1]["content"])

Chat models that fit the RTX 4070 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
4BQ8_0≈5.8 GBFits in memory~65 tok/sCommercial OK
3.2BQ8_0≈4.7 GBFits in memory~81 tok/sConditions
8BQ8_0≈9.8 GBFits in memory~32 tok/sConditions
3.8BQ8_0≈5.5 GBFits in memory~68 tok/sCommercial OK
4BQ8_0≈5.8 GBFits in memory~65 tok/sCommercial OK
7.6BQ8_0≈8.9 GBFits in memory~34 tok/sCommercial OK

Other GPUs that run Qwen3.8-4B-Distill

Frequently asked questions

Can the RTX 4070 run Qwen3.8-4B-Distill?
Yes. The RTX 4070 runs Qwen3.8-4B-Distill fully in VRAM at FP16 / BF16 (about 10.0 GB of 12 GB), at roughly 30 tokens/s.
How fast does Qwen3.8-4B-Distill run on the RTX 4070?
An estimate from memory bandwidth gives about 30 tokens/s at FP16 / BF16.
Which quantisation of Qwen3.8-4B-Distill should I use on the RTX 4070?
FP16 / BF16: the largest one that fits. Expected quality: lossless compared with the full-precision weights.

Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.