Qwen3.5 122B A10B needs ~87.8 GB VRAM. Gaudi 3 128GB has 128.0 GB. With Q3_K_M quantization, expect ~40 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
150.1 GB over capacity — needs offload or smaller quantization
Fit status
Too heavy
Decode
2.3 tok/s
TTFT
85025 ms
Safe context
4K
Memory
278.1 GB / 128.0 GB
Offload
50%
The raw memory story may look fine, but the software ecosystem is still a constraint here.
Runtime ecosystem is narrower than CUDA
Intel GPUs can look attractive on memory per dollar, but local AI tooling, kernels, and model coverage are still broader and easier on CUDA today.
Prefer CUDA if you want the path of least resistance
If your goal is maximum runtime coverage, easier troubleshooting, and better support for new local AI releases, CUDA is usually still the safer upgrade path.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | C | Runs well | 40.3 tok/s | 2621 ms | 61K |
| Coding | C | Runs well | 40.3 tok/s | 4805 ms | 61K |
| Agentic Coding | C | Runs well | 40.3 tok/s | 6989 ms | 61K |
| Reasoning | C | Runs well | 40.3 tok/s | 5678 ms | 61K |
| RAG | C | Runs well | 40.3 tok/s | 8736 ms | 61K |
Inference speed
Estimated decode speed (tokens/sec) for Qwen3.5 122B A10B at Q3_K_M across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is MacBook Pro M4 Max 128GB at ~9 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? |
|---|---|---|---|---|
MacBook Pro M4 Max 128GB | 128 GB | Q3_K_M | 9.3 | Offloads |
Mac Studio M3 Ultra 256GB | 256 GB | Q3_K_M | 8.7 | Fits |
Mac Studio M2 Ultra 128GB | 128 GB | Q3_K_M | 7.2 | Offloads |
Mac Studio M1 Ultra 128GB | 128 GB | Q3_K_M | 6.8 | Offloads |
2× RX 7900 XTX 24GB | 48 GB | Q3_K_M | 4.7 | Too big |
MacBook Pro M4 Max 64GB | 64 GB | Q3_K_M | 4.5 | Too big |
 RTX 5090 32GB | 32 GB | Q3_K_M | 2.8 | Too big |
MacBook Pro M4 Pro 48GB | 48 GB | Q3_K_M | 2.6 | Too big |
| 2× RTX 4090 24GB | 48 GB | Q3_K_M | 2.5 | Too big |
| 2× RTX 3090 24GB | 48 GB | Q3_K_M | 2.3 | Too big |
| RTX 4090 24GB | 24 GB | Q3_K_M | 2.0 | Too big |
| RTX 4080 Super 16GB | 16 GB | Q3_K_M | 2.0 | Too big |
| RTX 3090 24GB | 24 GB | Q3_K_M | 2.0 | Too big |
| RTX 4070 12GB | 12 GB | Q3_K_M | 2.0 | Too big |
| RTX 3060 12GB | 12 GB | Q3_K_M | 2.0 | Too big |
| RTX 4060 8GB | 8 GB | Q3_K_M | 2.0 | Too big |
RX 7900 XTX 24GB | 24 GB | Q3_K_M | 2.0 | Too big |
MacBook Pro M3 Max 64GB | 64 GB | Q3_K_M | 2.0 | Too big |
MacBook Pro M1 Max 64GB | 64 GB | Q3_K_M | 2.0 | Too big |
| 4× RTX 3060 12GB | 48 GB | Q3_K_M | 2.0 | Too big |
Estimates for single-stream decoding at Q3_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 Qwen3.5 122B A10B (122B params) fits at each quantization level on Gaudi 3 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 47.6 GB | Low | C45 |
Q3_K_S | 3 | 59.8 GB | Low | C47 |
NVFP4 | 4 | 68.3 GB | Medium | C48 |
Q4_K_M | 4 | 74.4 GB | Medium | C48 |
Q5_K_M | 5 | 87.8 GB | High | C48 |
Q6_KBest for your GPU | 6 | 100.0 GB | High | C48 |
Q8_0 | 8 | 130.5 GB | Very High | F0 |
F16 | 16 | 250.1 GB | Maximum | F0 |
Copy-paste commands to run Qwen3.5 122B A10B on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "unsloth/Qwen3.5-122B-A10B-GGUF" \
--hf-file "Qwen3.5-122B-A10B-GGUF-Q3_K_M.gguf" \
-c 4096 -ngl 99アップグレードオプション
NVIDIA H200 141GBコスパ重視Raises estimated decode speed by about 56%.
Moves you onto CUDA, which still has the broadest local-AI runtime coverage.
This is not only a hardware jump. It also gives you a cleaner runtime ecosystem for local LLM tooling.
〜$30,000 MSRP
NVIDIA H200 PCIe 141GBコスパ最良Raises estimated decode speed by about 56%.
Moves you onto CUDA, which still has the broadest local-AI runtime coverage.
This is not only a hardware jump. It also gives you a cleaner runtime ecosystem for local LLM tooling.
〜$30,000 MSRP
Yes, Gaudi 3 128GB can run Qwen3.5 122B A10B with a C grade (Runs well). Expected decode speed: 40.3 tok/s.
Qwen3.5 122B A10B (122B parameters) requires approximately 87.8 GB of memory with Q3_K_M quantization.
The recommended quantization for Qwen3.5 122B A10B is Q3_K_M, which balances quality and memory efficiency.
On Gaudi 3 128GB, Qwen3.5 122B A10B achieves approximately 40.3 tokens per second decode speed with a time-to-first-token of 4805ms using Q3_K_M quantization.
For coding workloads, Qwen3.5 122B A10B on Gaudi 3 128GB receives a C grade with 40.3 tok/s and 61K context.
On Gaudi 3 128GB, Qwen3.5 122B A10B can safely use up to 61K tokens of context. The model's official context limit is —, but available memory constrains the safe maximum.
Prefer CUDA if you want the path of least resistance. If your goal is maximum runtime coverage, easier troubleshooting, and better support for new local AI releases, CUDA is usually still the safer upgrade path.
Often yes, if your goal is the easiest setup and the widest runtime support. Intel can offer attractive memory capacity, but CUDA still tends to win on tooling maturity, guides, kernels, and model coverage for local AI.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/hf-unsloth--qwen3-5-122b-a10b-gguf-on-gaudi-3-128gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: