Codestral-22B on NVIDIA RTX 4090 (24GB VRAM)
Full FP16 (48.14 GB) exceeds 24 GB VRAM, but 4-bit GGUF Q4_K_M / AWQ requires 16.27 GB (13.86 GB weights + 1.75 GB KV cache at 8K context), delivering ~46.4 tokens/sec.
🎛️ Interactive VRAM & KV-Cache Tuner
📊 Real-Time Memory Footprint
16.27 GB / 24 GB🚀 Ready-to-Run Launch Command
ollama run codestral:22b # Or with llama.cpp server (8K context, Flash Attention): llama-server -hf mistralai/Codestral-22B-v0.1-GGUF -c 8192 -ngl 99 -fa
Need 16.27 GB+ VRAM to run Codestral-22B?
Recommended cloud setup for 4-bit: RTX 4090 (24GB) (or 48.14 GB for full FP16). Spin up a pre-configured vLLM / Ollama instance in 60 seconds:
| Provider | Best For | RTX 4090 (24GB) | A100 (80GB) | H100 (80GB) | Bonus / Trial | Action |
|---|---|---|---|---|---|---|
| RunPod | 1-Click vLLM / Ollama Pods | $0.44 / hr | $1.39 / hr | $2.39 / hr | Instant PyTorch / vLLM Pods + Serverless GPU | Deploy on RunPod → |
| Vast.ai | Cheapest Spot GPU Market | $0.29 / hr | $0.95 / hr | $1.79 / hr | Save 50–70% on RTX 4090 / 5090 / A100 rentals | Check Spot GPUs → |
| Vultr Cloud GPU | Enterprise SLA & Bare Metal | — | $1.79 / hr | $2.69 / hr | $300 Free Credit for New Users | Claim $300 Credit → |
| DigitalOcean GPU Droplets | Managed H100 Droplets | — | — | $2.99 / hr | $200 Free Trial (60 Days) | Get $200 Free → |
All Quantization Formats: Codestral-22B on NVIDIA RTX 4090
| Quantization Format | Bits/Weight | Weights VRAM | KV Cache (8K) | Total VRAM | Fits on 1x NVIDIA RTX 4090 (24GB)? | Est. Speed | GPUs Needed |
|---|---|---|---|---|---|---|---|
| FP16 / BF16 (Unquantized) | 16 bpw | 45.73 GB | 1.75 GB | 48.14 GB | No (Over by 24.1GB) | ~1.8 tok/s | 3x NVIDIA RTX 4090 |
| FP8 (W8A8 / E4M3) | 8.5 bpw | 24.3 GB | 1.75 GB | 26.71 GB | No (Over by 2.7GB) | ~3.4 tok/s | 2x NVIDIA RTX 4090 |
| GGUF Q8_0 (8-bit) | 8.5 bpw | 24.3 GB | 1.75 GB | 26.71 GB | No (Over by 2.7GB) | ~3.4 tok/s | 2x NVIDIA RTX 4090 |
| GGUF Q6_K (6-bit) | 6.56 bpw | 18.75 GB | 1.75 GB | 21.16 GB | Tight (88%) | ~34.3 tok/s | 1x NVIDIA RTX 4090 |
| GGUF Q5_K_M (5-bit) | 5.69 bpw | 16.26 GB | 1.75 GB | 18.67 GB | Yes (78%) | ~39.6 tok/s | 1x NVIDIA RTX 4090 |
| GGUF Q4_K_M / AWQ (4-bit) | 4.85 bpw | 13.86 GB | 1.75 GB | 16.27 GB | Yes (68%) | ~46.4 tok/s | 1x NVIDIA RTX 4090 |
| GGUF IQ4_XS (4.25 bpw) | 4.25 bpw | 12.15 GB | 1.75 GB | 14.56 GB | Yes (61%) | ~53 tok/s | 1x NVIDIA RTX 4090 |
| GGUF Q3_K_M (3-bit) | 3.91 bpw | 11.18 GB | 1.75 GB | 13.59 GB | Yes (57%) | ~57.6 tok/s | 1x NVIDIA RTX 4090 |
Frequently Asked Questions (Codestral-22B + NVIDIA RTX 4090)
Can NVIDIA RTX 4090 (24GB) run Codestral-22B locally?
Yes — NVIDIA RTX 4090 (24GB) can run Codestral-22B using GGUF Q6_K (6-bit). Full FP16 (48.14 GB) exceeds 24 GB VRAM, but 4-bit GGUF Q4_K_M / AWQ requires 16.27 GB (13.86 GB weights + 1.75 GB KV cache at 8K context), delivering ~46.4 tokens/sec.
How is KV Cache calculated for Codestral-22B?
Codestral-22B uses Dense-GQA with 56 transformer layers, 48 attention heads, and 8 KV heads (head dimension 128). At 8,192 tokens in FP16, KV cache consumes 1.75 GB. Enabling 8-bit KV cache (-ctk q8_0 -ctv q8_0 in llama.cpp or --kv-cache-dtype fp8 in vLLM) cuts this in half to 0.88 GB.