Can Qwen3.5 122B A10B run on AMD Instinct MI210 64GB?

Q: What should I upgrade first if Qwen3.5 122B A10B feels slow on AMD Instinct MI210 64GB?

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.

YES — With Q2_K

D38Poor

Estimated from fit model

Qwen3.5 122B A10B needs ~69.2 GB VRAM. AMD Instinct MI210 64GB has 64.0 GB. With Q2_K quantization, expect ~13 tok/s.

Runtime: llama.cppCapacity: OffloadBandwidth: HighStack: StandardBottleneck: Host offload

Operating mode

Choose the run profile you care about

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.

Qwen3.5 122B A10B at Q3_K_M needs 81.4 GB — too much for AMD Instinct MI210 64GB (64.0 GB). Runs at Q2_K (69.2 GB) with low quality.

Capabilities:

Select quantization to explore

F16 (Maximum quality) — 271.7 GB, exceeds 64.0 GB available

271.7 GB required64.0 GB available

425% VRAM needed

207.7 GB over capacity — needs offload or smaller quantization

Fit status

Too heavy

Decode

2.0 tok/s

TTFT

96800 ms

Safe context

Memory

271.7 GB / 64.0 GB

Offload

80%

Memory breakdown

Weights250.1 GB

KV Cache14.3 GB

Runtime0.9 GB

Headroom6.4 GB

See how fast it feels

With memory offload — actual speed may be lower

See how fast it feelsQwen3.5 122B A10B on AMD Instinct MI210 64GB
1st promptCold start — includes initialization
>What is local AI inference?
Local AI inference means running an AI language model directly on your own hardware — your laptop, desktop, or server — instead of sending requests to a remote cloud API.

When you run inference locally the model weights are loaded into your GPU or unified memory. Each token you generate requires reading those weights from memory, so memory bandwidth is the main bottleneck for decode speed.

Key benefits of running locally:
  - Full privacy: your prompts never leave your machine
  - No per-token cost or rate limits
  - Works offline once the model is downloaded
  - Latency depends only on your hardware
2nd promptWarm — model ready, no init overhead
>How much VRAM do I need?
It depends on the model size and quantization level. A rough rule of thumb:

  Model size     Q4 (4-bit)    Q8 (8-bit)    FP16
  7B params      ~4.3 GB       ~7.5 GB       ~14 GB
  13B params     ~7.9 GB       ~13.9 GB      ~26 GB
  70B params     ~42.7 GB      ~74.9 GB      ~140 GB

Most people use 4-bit quantization (Q4_K_M) which gives 90-95% of full quality at a fraction of the memory. A 24 GB GPU can comfortably run most 7B-13B models.
Estimated: 2.0 tok/s decode · 96.8s TTFT (warm) · 5 tok/s prefill

What limits this setup

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.

Best improvement path

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 3.6 GB of extra host RAM just for the offloaded portion, before OS and other tools.

Performance by workload

Workload	Grade	Fit	Decode	TTFT	Context
Chat	D	Very compromised (needs ~8.2 GB host RAM)	9.5 tok/s	11102 ms	4K
Coding	F	Too heavy	7.8 tok/s	24700 ms	4K
Agentic Coding	F	Too heavy	5.6 tok/s	50510 ms	4K
Reasoning	F	Too heavy	7.8 tok/s	29190 ms	4K
RAG	F	Too heavy	5.6 tok/s	63138 ms	4K

Inference speed

Qwen3.5 122B A10B inference speed — tokens per second by GPU & Mac

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

Quantization options

How Qwen3.5 122B A10B (122B params) fits at each quantization level on AMD Instinct MI210 64GB (64.0 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_KBest for your GPU	2	47.6 GB	Low	C48
Q3_K_S	3	59.8 GB	Low	F0

Get started

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

Upgrade options

Hardware that runs Qwen3.5 122B A10B well

AMD Instinct MI350X 288GBBudget pick

288 GB VRAM (+224)8000 GB/s (+6362)

Makes the model fit on the accelerator instead of staying completely out of reach.90.9 tok/s decode

Makes the model fit on the accelerator instead of staying completely out of reach.

Removes host-memory offload, which is usually the single biggest latency and throughput win.

~$8,000 MSRP

AMD Instinct MI300A 128GBBest value

128 GB VRAM (+64)5300 GB/s (+3662)

Makes the model fit on the accelerator instead of staying completely out of reach.57.7 tok/s decode

Makes the model fit on the accelerator instead of staying completely out of reach.

Removes host-memory offload, which is usually the single biggest latency and throughput win.

~$12,000 MSRP

AMD Instinct MI250X 128GBAMD upgrade

128 GB VRAM (+64)3200 GB/s (+1562)

Makes the model fit on the accelerator instead of staying completely out of reach.38.8 tok/s decode

Makes the model fit on the accelerator instead of staying completely out of reach.

Removes host-memory offload, which is usually the single biggest latency and throughput win.

~$15,000 MSRP

Frequently asked questions

Can AMD Instinct MI210 64GB run Qwen3.5 122B A10B?

Yes, AMD Instinct MI210 64GB can run Qwen3.5 122B A10B at Q2_K quantization (Very compromised (needs ~3.6 GB host RAM)). The recommended Q3_K_M requires 81.4 GB which exceeds available memory, but at Q2_K it needs only 69.2 GB. Expected decode speed: 12.7 tok/s.

How much VRAM does Qwen3.5 122B A10B need?

Qwen3.5 122B A10B (122B parameters) requires approximately 81.4 GB at Q3_K_M quantization. On AMD Instinct MI210 64GB, it fits at Q2_K using 69.2 GB.

What is the best quantization for Qwen3.5 122B A10B?

The recommended quantization is Q3_K_M, but on AMD Instinct MI210 64GB the best fitting quantization is Q2_K, which uses 69.2 GB.

What speed will Qwen3.5 122B A10B run at on AMD Instinct MI210 64GB?

On AMD Instinct MI210 64GB, Qwen3.5 122B A10B achieves approximately 12.7 tokens per second decode speed with a time-to-first-token of 15280ms using Q2_K quantization.

Can AMD Instinct MI210 64GB run Qwen3.5 122B A10B for coding?

For coding workloads, Qwen3.5 122B A10B on AMD Instinct MI210 64GB receives a F grade with 7.8 tok/s and 4K context.

What context window can Qwen3.5 122B A10B use on AMD Instinct MI210 64GB?

On AMD Instinct MI210 64GB, Qwen3.5 122B A10B can safely use up to 10K tokens of context at Q2_K quantization. The model's official context limit is —, but available memory constrains the safe maximum.

See all results for AMD Instinct MI210 64GB See all hardware for Qwen3.5 122B A10B

Embed this result▼

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-instinct-mi210-64gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>

Preview: