Running 70B Models Locally: Multi-GPU, CPU Offload, and What Actually Works

70B parameter models are the threshold between "good" and "GPT-4 competitive." Running them locally is completely achievable in 2026 — but requires understanding your options. This guide covers every realistic path from budget to performance-first.

Memory Requirements (70B at Different Quantizations)

QuantizationModel SizeHardwareSpeed
Q2_K~25 GBSingle RTX 4090 (24GB+RAM)~8 tok/s
Q3_K_M~31 GBSingle RTX 4090 + CPU offload~5 tok/s
Q4_K_M~42 GB2x RTX 4090 or Mac Ultra~22 tok/s ★
Q5_K_M~50 GBMac Ultra M4 or 3x RTX 4090~18 tok/s
Q8_0~74 GB4x RTX 4090 or 8x P40~12 tok/s

Option 1: Single RTX 4090 with CPU Offload

You can run Llama 3.3 70B or DeepSeek-R1-70B on a single RTX 4090 by offloading some layers to CPU RAM. Requires 64GB+ system RAM.

./build/bin/llama-server \
  -m ./models/llama3.3-70b-q4_k_m.gguf \
  --n-gpu-layers 56 \    # offload 56/80 layers to GPU, rest to CPU RAM
  --ctx-size 8192 \
  --batch-size 512 \
  --host 0.0.0.0 --port 8080

# Tune n-gpu-layers: more layers = faster but needs more VRAM
# RTX 4090 (24GB) can hold ~56-58 layers of 70B Q4

Option 2: Two RTX 4090s (Recommended for Performance)

Two RTX 4090s give you 48GB combined VRAM — enough for 70B at Q4 fully on GPU. For hardware recommendations, LLMHardware.io's dual-GPU build guide has tested configurations.

./build/bin/llama-server \
  -m ./models/deepseek-r1-70b-q4_k_m.gguf \
  --n-gpu-layers 999 \
  --tensor-split 1,1 \   # equal split between GPU 0 and 1
  --flash-attn \
  --ctx-size 32768 \
  --host 0.0.0.0 --port 8080

Option 3: Mac Studio / Mac Pro (Apple Silicon)

The Mac Studio M4 Ultra with 192GB unified memory is arguably the best single-unit 70B runner available. Unified memory eliminates the VRAM constraint entirely:

ollama pull llama3.3:70b-q5_K_M
ollama run llama3.3:70b-q5_K_M

# Or via llama.cpp with Metal acceleration:
./build/bin/llama-server \
  -m ./models/llama3.3-70b-q5_k_m.gguf \
  --n-gpu-layers 999 \   # all layers to Metal GPU
  --ctx-size 65536 \
  --flash-attn

Speed Comparison: 70B Q4_K_M

SetupTokens/secCost (2026)
RTX 4090 + CPU offload~5 tok/s$1,100 + $600+ for 64GB RAM
2x RTX 4090~22 tok/s~$2,200
Mac Studio M4 Max~15 tok/s$3,199
Mac Studio M4 Ultra~28 tok/s$4,999

Comments

Post a Comment

Popular posts from this blog

AMD RX 9070 XT for Local LLMs: ROCm Performance Guide (2026)

The AMD RX 9070 XT launched in early 2026 at $599 with 16GB GDDR6 and 648 GB/s memory bandwidth. For local LLM users who want a mid-range GPU without NVIDIA prices, this changes the calculus significantly. Here is what it actually delivers. RX 9070 XT Specs vs Competition GPU VRAM Bandwidth Price LLM Score RX 9070 XT 16 GB 648 GB/s $599 ~60 tok/s RTX 4070 Ti Super 16 GB 672 GB/s $799 ~62 tok/s RTX 4080 Super 16 GB 736 GB/s $999 ~68 tok/s RTX 4090 24 GB 1,008 GB/s ~$1,100 ~90 tok/s At $599 with 16GB VRAM, the RX 9070 XT is the best value GPU for local LLMs in the mid-range tier in 2026. See the full comparison at LLMHardware.io's AMD GPU guide . ROCm Setup on Linux (2026) # Ubuntu 22.04/24.04 sudo apt install rocm-hip-libraries hip-base # or use the AMD installer: wget https://repo.radeon.com/amdgpu-install/6.3/ubuntu/jammy/amdgpu-install_6.3.60300-1_all.deb sudo dpkg -i amdgpu-install*.deb sudo amdgpu-install --usecase=rocm,lrt,openclsdk # Verify ROCm sees your GPU r...
Read more

llama.cpp flash attention: how to enable and benchmark

Image
Flash attention computes attention with far less memory traffic. The math is identical; the memory movement is not. In llama.cpp it matters most in two situations: long context windows and tight video memory (VRAM). With flash attention enabled, the key-value cache shrinks substantially. A 32k-token context window becomes realistic on a 16GB card, and the prompt-processing phase speeds up, so long documents get ingested faster. The catch is compatibility: not every model and quantization combination supports it, and on some setups the speedup is modest. If you run short chats on a strong GPU, you may barely notice. Where it shines is retrieval-augmented generation (RAG) and long-document work on mid-range cards. Enable it, compare your tokens per second, and keep it if your workload benefits. For the common case of pasting a long PDF and asking questions, it is close to free performance. See how context length changes VRAM needs .
Read more

SillyTavern with Local LLMs: Complete Setup for AI Companions

Image
SillyTavern is the most feature-rich frontend for character-based AI conversations. What began as an AI roleplay tool has evolved into a powerful platform with persona management, memory systems, extensive prompt engineering tools, and connections to every major local inference backend. Here's how to set it up with local models. Quick connect: SillyTavern connects to Ollama, KoboldCpp, llama.cpp server, and TabbyAPI (ExLlamaV2) out of the box. Any of the hardware setups at LLMHardware.io provides a solid backend for SillyTavern. Installation Install via git: git clone https://github.com/SillyTavern/SillyTavern && cd SillyTavern && npm install then node server.js . Access at localhost:8000. For persistent installation, use the included start.sh script. SillyTavern uses Node.js and runs on any platform. The companion app (STMP) enables mobile access to your SillyTavern instance via Tailscale or local network. Character Cards and the TavernAI Format Chara...
Read more