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.

Axion1-0 logo

Can the RTX 3090 run Axion-1.0-35B?

Axion-1.0-35B·35.1B (3.9B active)·GeForce RTX 3090·24 GB VRAM

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

Yes. The RTX 3090 runs Axion-1.0-35B fully in VRAM at Q4_K_M (about 22 GB of 24 GB), at roughly 218 tokens/s.

best quant
Q4_K_M
needs
22 GB
usable
24 GB
est. speed
~218 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the Axion-1.0-35B page →

Every quantisation of Axion-1.0-35B on the RTX 3090

QuantQualityMemoryFitSpeed
Q2_K Heavy loss14 GBfits~401 tok/s
Q3_K_M Noticeable loss18 GBfits~270 tok/s
Q4_K_M pickSmall loss22 GBfits~218 tok/s
Q5_K_M Near-lossless25 GBoffload~106 tok/s
Q6_K Near-lossless29 GBoffload~45 tok/s
Q8_0 Lossless37 GBoffload~20 tok/s
FP16 / BF16 Lossless70 GBtoo big—

Speed is a bandwidth-based estimate for generating text (936 GB/s); prompt processing and long contexts are slower.

Measured speed on the RTX 3090

No community reports for this pair yet.

Report your own speed →

Run Axion-1.0-35B on the RTX 3090

Using Q4_K_M, the best fit for RTX 3090.

Ollama · Q4_K_M straight from Hugging Facedocs
ollama run hf.co/mradermacher/Axion-1.0-35B-GGUF:Q4_K_M
llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf mradermacher/Axion-1.0-35B-GGUF:Q4_K_M

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 mradermacher/Axion-1.0-35B-GGUF@Q4_K_M

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

vLLM · Serve an OpenAI-compatible APIdocs
vllm serve Axion1-0/Axion-1.0-35B
vLLM · Docker Compose (NVIDIA)docs
services:
  vllm:
    image: vllm/vllm-openai:latest
    command: Axion1-0/Axion-1.0-35B
    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 Axion1-0/Axion-1.0-35B --host 0.0.0.0 --port 30000
SGLang · Docker Compose (NVIDIA)docs
services:
  sglang:
    image: lmsysorg/sglang:latest
    command: sglang serve Axion1-0/Axion-1.0-35B --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="Axion1-0/Axion-1.0-35B", 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 3090 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
36BA3BQ4_K_M≈23 GBFits in memory~283 tok/sCommercial OK
32.8BQ4_K_M≈22 GBFits in memory~26 tok/sCommercial OK
30.5BA3BQ5_K_M≈22 GBFits in memory~241 tok/sCommercial OK
20.9BA4.2BQ8_0≈13 GBFits in memory~116 tok/sCommercial OK
32.8BQ4_K_M≈22 GBFits in memory~26 tok/sCommercial OK
33.4BA2.2BQ4_K_M≈22 GBFits in memory~380 tok/sCommercial OK

Other GPUs that run Axion-1.0-35B

Frequently asked questions

Can the RTX 3090 run Axion-1.0-35B?
Yes. The RTX 3090 runs Axion-1.0-35B fully in VRAM at Q4_K_M (about 22 GB of 24 GB), at roughly 218 tokens/s.
How fast does Axion-1.0-35B run on the RTX 3090?
An estimate from memory bandwidth gives about 218 tokens/s at Q4_K_M.
Which quantisation of Axion-1.0-35B should I use on the RTX 3090?
Q4_K_M: the largest one that fits. Expected quality: small loss 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.