Post
45
Nanbeige4.3-33B and Qwen3.8-27B are close in parameter scale, but their published decoders have 52 versus 16 full-attention layers.
We added a context slider to alphaBell Model Observatory. For one 32,768-token sequence, raw f16 full-attention K/V would be 6.98 GB versus 2.15 GB if a runtime enables caching. This is a 3.25× difference for that tensor component. Nanbeige's released model and generation configs set use_cache:false; Qwen recurrent state, weights and runtime overhead are separate.
Explore the layer map and share a context length: https://alphabell-model-observatory.static.hf.space/architecture/index.html?tokens=32768#kv-title
Source-pinned writeup: https://huggingface.co/blog/alphabell/nanbeige-qwen-layer-map
If you have a matched speed run, please share the runtime build, quantization, KV dtype, prompt/output lengths and separate prefill/decode rates.
Space: alphabell/Model-Observatory
We added a context slider to alphaBell Model Observatory. For one 32,768-token sequence, raw f16 full-attention K/V would be 6.98 GB versus 2.15 GB if a runtime enables caching. This is a 3.25× difference for that tensor component. Nanbeige's released model and generation configs set use_cache:false; Qwen recurrent state, weights and runtime overhead are separate.
Explore the layer map and share a context length: https://alphabell-model-observatory.static.hf.space/architecture/index.html?tokens=32768#kv-title
Source-pinned writeup: https://huggingface.co/blog/alphabell/nanbeige-qwen-layer-map
If you have a matched speed run, please share the runtime build, quantization, KV dtype, prompt/output lengths and separate prefill/decode rates.
Space: alphabell/Model-Observatory