Qwen2.5 · Apache-2.0 · HF: Qwen/Qwen2.5-Coder-32B-Instruct
Qwen2.5 Coder 32B — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.66 in / $1.00 out per 1M
Parameters
32.8B Dense
Attention
GQA (8 KV / 40 Q)
Layers × Hidden
64 × 5120
Max Native Context
32k tokens
MoE Experts
N/A (Dense)
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 62.32 GiB | 65.77 GiB | 71.77 GiB | Lossless Reference |
| Q8_0 | 8.50 | 33.11 GiB | 36.56 GiB | 42.56 GiB | Near-Lossless |
| Q6_K | 6.57 | 25.59 GiB | 29.04 GiB | 35.04 GiB | High Precision |
| Q5_K_M | 5.69 | 22.16 GiB | 25.61 GiB | 31.61 GiB | Balanced High |
| Q4_K_M | 4.85 | 18.89 GiB | 22.34 GiB | 28.34 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 17.53 GiB | 20.98 GiB | 26.98 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 16.55 GiB | 20.01 GiB | 26.01 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 15.23 GiB | 18.68 GiB | 24.68 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 28.3 / 140.4 GiB | 139.7 tok/s | 0.43s | 32k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 28.3 / 79.4 GiB | 97.5 tok/s | 0.43s | 32k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 28.3 / 79.4 GiB | 57.9 tok/s | 1.35s | 32k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 28.3 / 95.0 GiB | 52.1 tok/s | 1.68s | 32k | Simulate → |
| Cloud NVIDIA L40S (48 GB) | COMFORTABLE | 28.3 / 47.4 GiB | 24.5 tok/s | 1.21s | 32k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 28.1 / 199.7 GiB | 22.1 tok/s | 8.35s | 32k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 28.1 / 399.4 GiB | 22.1 tok/s | 8.35s | 32k | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | COMFORTABLE | 28.3 / 46.6 GiB | 21.8 tok/s | 7.85s | 32k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 28.1 / 149.8 GiB | 21.6 tok/s | 9.22s | 32k | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | COMFORTABLE | 28.1 / 49.9 GiB | 14.6 tok/s | 12.70s | 32k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 28.1 / 99.8 GiB | 14.6 tok/s | 12.70s | 32k | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | COMFORTABLE | 28.1 / 49.9 GiB | 10.5 tok/s | 17.02s | 32k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 28.1 / 99.8 GiB | 10.5 tok/s | 17.02s | 32k | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | COMFORTABLE | 28.1 / 37.4 GiB | 7.0 tok/s | 26.24s | 32k | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | COMFORTABLE | 28.1 / 49.9 GiB | 7.0 tok/s | 26.24s | 32k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 28.1 / 96.0 GiB | 6.6 tok/s | 9.45s | 32k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | COMFORTABLE | 28.1 / 48.0 GiB | 6.6 tok/s | 9.45s | 32k | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | COMFORTABLE | 28.1 / 48.0 GiB | 2.9 tok/s | 27.06s | 32k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | TIGHT | 28.3 / 31.0 GiB | 50.9 tok/s | 2.09s | 32k | Simulate → |
| NVIDIA RTX 4090 (24 GB) | OOM | 28.3 / 23.0 GiB | OOM | — | 11k | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 28.3 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | OOM | 28.3 / 23.0 GiB | OOM | — | 11k | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 28.3 / 15.0 GiB | OOM | — | OOM | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | OOM | 28.3 / 23.0 GiB | OOM | — | 11k | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 28.3 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 28.3 / 11.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 28.1 / 10.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (24 GB) | OOM | 28.1 / 18.0 GiB | OOM | — | OOM | Simulate → |
Interactive Qwen2.5 Coder 32B Simulator
32.8B Dense · GQA
1792 GB/s · 250 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 32k
Memory Allocation Footprint 28.34 / 95.00 GiB usable
Weights: 18.89 GiB
KV Cache: 8.00 GiB
Overhead: 1.45 GiB
Headroom: 66.66 GiB
Decode Speed
52.1
tok/s (batch=1)
Prefill Speed
2490
prompt tok/s
TTFT Latency
1.68s
for 4,096 tok
Active Params
32.8B
of 32.8B total
llama.cpp Server Command
llama-server -hf Qwen/Qwen2.5-Coder-32B-Instruct:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 1.00 GiB | 20.99 GiB | 70.5 tok/s | COMFORTABLE |
| 8k | 2.00 GiB | 22.34 GiB | 67.1 tok/s | COMFORTABLE |
| 16k | 4.00 GiB | 24.34 GiB | 61.3 tok/s | COMFORTABLE |
| 32k | 8.00 GiB | 28.34 GiB | 52.1 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$167.91
85.5h active + power
Cloud GPU/mo
$181.88
@$1.85/hr spot/pod
Serverless API/mo
$48.00
Payback: 164.7 mo
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — Qwen2.5 Coder 32B
How much VRAM does Qwen2.5 Coder 32B require in Q4_K_M?
Qwen2.5 Coder 32B weights occupy 18.89 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 28.34 GiB.
What is the cheapest local hardware that runs Qwen2.5 Coder 32B?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs Qwen2.5 Coder 32B Q4_K_M at ~52.1 tok/s decode and 1.68s TTFT.
How does Qwen2.5 Coder 32B scale KV cache at long context?
Qwen2.5 Coder 32B uses GQA with 8 KV heads across 64 layers (128 head_dim).