What GPU do I need to run Qwen/Qwen3.5-122B-A10B-FP8?
125.1B parameters, published in F8_E4M3. View on Hugging Face
Set up Qwen3.5-122B-A10B-FP8Qwen3.5-122B-A10B-FP8 is published by Qwen on Hugging Face, with 1,786,023 downloads and 115 likes to date. It's a Qwen3_5MoeForConditionalGeneration model built for image-text-to-text, published natively in F8_E4M3.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| FP8 (native) | 116.5 GB | 139.8 GB | AMD MI300X | 1 | $2.63/hr |
| cheaper alt. | RTX 4000 SFF Ada | 7 | $1.39/hr | ||
| INT4 (quantized) | 58.3 GB | 69.9 GB | A100 | 1 | $1.10/hr |
| cheaper alt. | RTX A5000 | 3 | $0.528/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run Qwen3.5-122B-A10B-FP8 at its published (F8_E4M3) precision: 1× AMD MI300X, at $2.63/hr per GPU ($2.63/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
Qwen3.5-122B-A10B-FP8: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Hybrid: only the full-attention layers cache per token; the other layers keep a fixed-size state. Cached per token: 24,576 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.09 GB |
| 32K | 0.75 GB |
| 128K | 3.0 GB |
| 256K (model maximum) | 6.0 GB |
Computed from the layer, head and window counts in the model's own configuration (Qwen/Qwen3.5-122B-A10B-FP8), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
Qwen3.5-122B-A10B-FP8: common questions
Can Qwen3.5-122B-A10B-FP8 run on a single GPU?
Yes, but not on a desktop card. At FP8 (native) it needs 139.8 GB of VRAM, more than a single 32 GB desktop card holds. The cheapest live single-card fit is a 192.0 GB AMD MI300X at $2.63/hr.
Is Qwen3.5-122B-A10B-FP8 already quantized?
Yes. It is published in FP8, one byte per parameter, so the 139.8 GB figure is already a quantized footprint rather than a full-precision one. Only INT4 goes below it, at 69.9 GB. A GPU without FP8 tensor cores cannot run it as published, which is why cards here are matched on precision support and not on VRAM alone.
Does quantizing Qwen3.5-122B-A10B-FP8 lower the GPU bill?
Yes. At FP8 (native) the cheapest live fit is one AMD MI300X at $2.63/hr. At INT4 (quantized) it drops to one A100 at $1.10/hr, provided a quantized checkpoint exists for it.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Qwen3.5 models
- Qwen3.5-122B-A10B (125.1B, BF16)
- Qwen3.5-397B-A17B (403.4B, BF16)
- Qwen3.5-397B-A17B-FP8 (403.4B, F8_E4M3)
- Qwen3.5-35B-A3B-FP8 (36.0B, F8_E4M3)
- Qwen3.5-35B-A3B (36.0B, BF16)
Lineage
Based on Qwen3.5-122B-A10B.
Alternatives at this size
Other models for image-text-to-text within about a third of Qwen3.5-122B-A10B-FP8's 125.1B parameters, from other model lines.
- Llama-4-Scout-17B-16E-Instruct (108.6B, BF16)
- GLM-4.5V (107.7B, BF16)
- Darwin-180B-RSI (180.0B, BF16)
More on Qwen3.5-122B-A10B-FP8
Fits on a 141 GB GPU at FP8: every model that fits in 141 GB.
Fits on an 80 GB GPU at INT4: every model that fits in 80 GB.
Related reading: AMD MI300X pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
