From the model card
Fields below are copied from the tags and counters on the HuggingFace repository Intel/Qwen3.5-122B-A10B-int4-AutoRound at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- Intel
- Weight formats
- safetensors
- Lineage
-
- base model Qwen/Qwen3.5-122B-A10B
- quantized from Qwen/Qwen3.5-122B-A10B
- Papers cited
- arXiv:2309.05516
- Downloads (HF counter at last fetch)
- 358,065
- Likes (HF counter at last fetch)
- 41
- Model card
- https://huggingface.co/Intel/Qwen3.5-122B-A10B-int4-AutoRound
Use cases
- Deploying Qwen3.5-122B on Intel Gaudi accelerators
- Reducing memory footprint of a 122B MoE model for data-center inference
- Evaluating AutoRound quantization quality against GPTQ/AWQ on large MoE models
- Enterprise on-premise LLM serving on Intel infrastructure
Pros
- AutoRound typically outperforms GPTQ on complex benchmarks at INT4
- 122B scale with MoE efficiency (10B active) gives frontier-adjacent quality
- Intel-maintained quantization with published methodology
- A10B active parameters keeps per-token compute reasonable
Cons
- Optimized for Intel hardware — performance on NVIDIA GPUs may be suboptimal
- INT4 precision causes measurable accuracy drops on math and code vs BF16
- 122B total parameters require substantial storage and loading time
- AutoRound tooling is less widely adopted than GPTQ/AWQ in community infrastructure
Tags
safetensorsqwen3_5_moearxiv:2309.05516base_model:Qwen/Qwen3.5-122B-A10Bbase_model:quantized:Qwen/Qwen3.5-122B-A10B4-bitauto-roundregion:us