DeepSeek V2.5 236B needs ~222.7 GB VRAM. NVIDIA GB200 192GB has 192.0 GB. With Q4_K_M quantization, expect ~84 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
30.7 GB over capacity — needs offload or smaller quantization
Fit status
Very compromised (needs ~19.8 GB host RAM)
Decode
84.0 tok/s
TTFT
2305 ms
Safe context
8K
Memory
222.7 GB / 192.0 GB
Offload
10%
It fits through host-memory offload, and offload is the main reason performance drops.
CPU or host-memory offload is active
About 10% 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.
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 19.8 GB of extra host RAM just for the offloaded portion, before OS and other tools.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | S | Runs with offload (needs ~1 GB host RAM) | 106.0 tok/s | 996 ms | 8K |
| Coding | A | Very compromised (needs ~19.8 GB host RAM) | 84.0 tok/s | 2305 ms | 8K |
| Agentic Coding | F | Too heavy | 57.1 tok/s | 4929 ms | 8K |
| Reasoning | A | Very compromised (needs ~19.8 GB host RAM) | 84.0 tok/s | 2724 ms | 8K |
| RAG | F | Too heavy | 57.1 tok/s | 6161 ms | 8K |
Inference speed
Estimated decode speed (tokens/sec) for DeepSeek V2.5 236B at Q4_K_M across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is Mac Studio M3 Ultra 256GB at ~12 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 11.7 | Too big |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 6.1 | Too big |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 5.7 | Too big |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 4.5 | Too big |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 4.5 | Too big |
 RTX 5090 32GB | 32 GB | Q4_K_M | 3.4 | Too big |
2× RX 7900 XTX 24GB | 48 GB | Q4_K_M | 3.3 | Too big |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 3.1 | Too big |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 2.9 | Too big |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 2.7 | Too big |
| 2× RTX 4090 24GB | 48 GB | Q4_K_M | 2.3 | Too big |
| RTX 4090 24GB | 24 GB | Q4_K_M | 2.2 | Too big |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 2.0 | Too big |
| RTX 3090 24GB | 24 GB | Q4_K_M | 2.0 | Too big |
| RTX 4070 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 3060 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 4060 8GB | 8 GB | Q4_K_M | 2.0 | Too big |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 2.0 | Too big |
| 2× RTX 3090 24GB | 48 GB | Q4_K_M | 2.0 | Too big |
| 4× RTX 3060 12GB | 48 GB | Q4_K_M | 2.0 | Too big |
Estimates for single-stream decoding at Q4_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
How DeepSeek V2.5 236B (236B params) fits at each quantization level on NVIDIA GB200 192GB (192.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 92.0 GB | Low | A81 |
Q3_K_S | 3 | 115.6 GB | Low | A82 |
NVFP4 | 4 | 132.2 GB | Medium | A82 |
Q4_K_MBest for your GPU | 4 | 144.0 GB | Medium | A82 |
Q5_K_M | 5 | 169.9 GB | High | F0 |
Q6_K | 6 | 193.5 GB | High | F0 |
Q8_0 | 8 | 252.5 GB | Very High | F0 |
F16 | 16 | 483.8 GB | Maximum | F0 |
Copy-paste commands to run DeepSeek V2.5 236B on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "deepseek-ai/DeepSeek-V2.5" \
--hf-file "DeepSeek-V2.5-Q4_K_M.gguf" \
-c 4096 -ngl 99Your hardware
| Model | Params | Grade | Decode | Capabilities |
|---|---|---|---|---|
 DeepSeek V4 Flash | 284B | S | 144.8 tok/s |
Yes, NVIDIA GB200 192GB can run DeepSeek V2.5 236B with a A grade (Very compromised (needs ~19.8 GB host RAM)). Expected decode speed: 84.0 tok/s.
DeepSeek V2.5 236B (236B parameters) requires approximately 222.7 GB of memory with Q4_K_M quantization.
The recommended quantization for DeepSeek V2.5 236B is Q4_K_M, which balances quality and memory efficiency.
On NVIDIA GB200 192GB, DeepSeek V2.5 236B achieves approximately 84.0 tokens per second decode speed with a time-to-first-token of 2305ms using Q4_K_M quantization.
For coding workloads, DeepSeek V2.5 236B on NVIDIA GB200 192GB receives a A grade with 84.0 tok/s and 8K context.
On NVIDIA GB200 192GB, DeepSeek V2.5 236B can safely use up to 8K tokens of context. The model's official context limit is 131K, 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/deepseek-v2.5-236b-on-gb200-192gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: