Qwen3.5 35B A3B needs ~28.8 GB VRAM. Tesla P40 24GB has 24.0 GB. With Q4_K_M quantization, expect ~5 tok/s.
Operating mode
Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
Select quantization to explore
4.8 GB over capacity — needs offload or smaller quantization
Fit status
Very compromised (needs ~3.5 GB host RAM)
Decode
4.7 tok/s
TTFT
41274 ms
Safe context
4K
Memory
28.8 GB / 24.0 GB
Offload
20%
It fits through host-memory offload, and offload is the main reason performance drops.
CPU or host-memory offload is active
About 20% of the working set spills out of accelerator memory, which usually hurts latency and sustained decode throughput.
Very little memory headroom
You can run the model, but there is not much room left for longer context, bigger batches, extra apps, or future model updates.
Older PCIe generation
PCIe 3.0 is workable, but it compounds the penalty when you offload heavily or try to scale across multiple cards.
Remove offload with more accelerator memory
Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Buy headroom, not only minimum fit
A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.
Increase host RAM if you keep offloading
This setup may need roughly 3.5 GB of extra host RAM just for the offloaded portion, before OS and other tools.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | D | Very compromised (needs ~2.2 GB host RAM) | 5.5 tok/s | 19137 ms | 4K |
| Coding | D | Very compromised (needs ~3.5 GB host RAM) | 4.7 tok/s | 41274 ms | 4K |
| Agentic Coding | F | Too heavy | 3.5 tok/s | 80462 ms | 4K |
| Reasoning | D | Very compromised (needs ~3.5 GB host RAM) | 4.7 tok/s | 48779 ms | 4K |
| RAG | F | Too heavy | 3.5 tok/s | 100577 ms | 4K |
How Qwen3.5 35B A3B (35B params) fits at each quantization level on Tesla P40 24GB (24.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 13.7 GB | Low | C50 |
Q3_K_SBest for your GPU | 3 | 17.2 GB | Low | C50 |
NVFP4 | 4 | 19.6 GB | Medium | F0 |
Q4_K_M | 4 | 21.3 GB | Medium | F0 |
Q5_K_M | 5 | 25.2 GB | High | F0 |
Q6_K | 6 | 28.7 GB | High | F0 |
Q8_0 | 8 | 37.5 GB | Very High | F0 |
F16 | 16 | 71.8 GB | Maximum | F0 |
Copy-paste commands to run Qwen3.5 35B A3B on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "lmstudio-community/Qwen3.5-35B-A3B-GGUF" \
--hf-file "Qwen3.5-35B-A3B-GGUF-Q4_K_M.gguf" \
-c 4096 -ngl 99アップグレードオプション
RTX 5090 32GBコスパ重視Removes host-memory offload, which is usually the single biggest latency and throughput win.
Raises estimated decode speed by about 674%.
〜$1,999 MSRP
RTX PRO 4500 Blackwell 32GBコスパ最良Removes host-memory offload, which is usually the single biggest latency and throughput win.
Raises estimated decode speed by about 651%.
〜$2,499 MSRP
RTX 5000 Ada 32GBNVIDIAアップグレードRemoves host-memory offload, which is usually the single biggest latency and throughput win.
Raises estimated decode speed by about 360%.
〜$4,000 MSRP
Yes, Tesla P40 24GB can run Qwen3.5 35B A3B with a D grade (Very compromised (needs ~3.5 GB host RAM)). Expected decode speed: 4.7 tok/s.
Qwen3.5 35B A3B (35B parameters) requires approximately 28.8 GB of memory with Q4_K_M quantization.
The recommended quantization for Qwen3.5 35B A3B is Q4_K_M, which balances quality and memory efficiency.
On Tesla P40 24GB, Qwen3.5 35B A3B achieves approximately 4.7 tokens per second decode speed with a time-to-first-token of 41274ms using Q4_K_M quantization.
For coding workloads, Qwen3.5 35B A3B on Tesla P40 24GB receives a D grade with 4.7 tok/s and 4K context.
On Tesla P40 24GB, Qwen3.5 35B A3B can safely use up to 4K tokens of context. The model's official context limit is —, but available memory constrains the safe maximum.
Remove offload with more accelerator memory. Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/hf-lmstudio-community--qwen3-5-35b-a3b-gguf-on-tesla-p40-24gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: