Posting the link is enough: X, Reddit and Discord turn it into the card automatically.
Qwen3.8 Flash Next
Quant IQ1_S (UD)Params 177B (3B active)Size 72.5 GBArch qwen4exp A3B
on 2× AMD Mi50 32GB · ROCm
Generation
31.67tok/s
over 512 tokens
Prompt
366.79tok/s
over 2048 tokens
Model
- Quant
- IQ1_S (UD)
- Parameters
- 177B (3B active)
- File size
- 72.5 GB
- Architecture
- qwen4exp A3B
- File
- Qwen3.8-Flash-Next-UD-IQ1_S-00001-of-00003.gguf
Machine
- Devices
- 2× AMD Instinct MI60 / MI50
- Backend
- ROCm
- CPU
- Intel(R) Xeon(R) CPU E5-1650 v3 @ 3.50GHz
How it ran
- Engine
- llama.cpp
- Build
- b10811 (4e6a72bda)
- KV cache
- q8_0
- Flash attention
- on
- GPU layers
- all
- Batch / ubatch
- 4096 / 2048
- Threads
- 6
- Split mode
- layer
- Tensor split
- 0.00
- Context depth
- 0 (empty context)
Benched by webzone · Aug 30 · run at 2026-08-30 20:28:13 UTC
The benchmark exactly as it was pasted
ggml_cuda_init: found 2 ROCm devices (Total VRAM: 65504 MiB):
Device 0: AMD Instinct MI60 / MI50, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 32752 MiB
Device 1: AMD Instinct MI60 / MI50, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 32752 MiB
[
{
"build_commit": "4e6a72bda",
"build_number": 10811,
"cpu_info": "Intel(R) Xeon(R) CPU E5-1650 v3 @ 3.50GHz",
"gpu_info": "AMD Instinct MI60 / MI50, AMD Instinct MI60 / MI50",
"backends": "ROCm",
"model_filename": "/models/UD-IQ1_S/Qwen3.8-Flash-Next-UD-IQ1_S-00001-of-00003.gguf",
"model_type": "qwen4exp A3B IQ1_S - 1.5625 bpw",
"model_size": 72535436800, "model_n_params": 176943899520,
"n_batch": 4096,
"n_ubatch": 2048,
"n_threads": 6,
"cpu_mask": "0x0",
"cpu_strict": false,
"poll": 50,
"type_k": "q8_0",
"type_v": "q8_0",
"n_gpu_layers": 999,
"n_cpu_moe": 0,
"split_mode": "layer",
"tensor_parallel_size": 0,
"main_gpu": 0,
"no_kv_offload": false,
"flash_attn": 1,
"devices": "auto",
"tensor_split": "0.00",
"tensor_buft_overrides": "per_layer_token_embd=CPU",
"load_mode": "none",
"lazy_mode": "off",
"embeddings": false,
"no_op_offload": 0,
"no_host": false,
"no_repack": false,
"fit_target": 0,
"fit_min_ctx": 0,
"n_prompt": 2048,
"n_gen": 0,
"n_depth": 0,
"test_time": "2026-08-30T20:28:02Z",
"avg_ns": 5583603950,
"stddev_ns": 0,
"avg_ts": 366.788192,
"stddev_ts": 0.000000,
"samples_ns": [ 5583603950 ],
"samples_ts": [ 366.788 ]
},
{
"build_commit": "4e6a72bda",
"build_number": 10811,
"cpu_info": "Intel(R) Xeon(R) CPU E5-1650 v3 @ 3.50GHz",
"gpu_info": "AMD Instinct MI60 / MI50, AMD Instinct MI60 / MI50",
"backends": "ROCm",
"model_filename": "/models/UD-IQ1_S/Qwen3.8-Flash-Next-UD-IQ1_S-00001-of-00003.gguf", "model_type": "qwen4exp A3B IQ1_S - 1.5625 bpw",
"model_size": 72535436800,
"model_n_params": 176943899520,
"n_batch": 4096,
"n_ubatch": 2048,
"n_threads": 6,
"cpu_mask": "0x0",
"cpu_strict": false,
"poll": 50,
"type_k": "q8_0",
"type_v": "q8_0",
"n_gpu_layers": 999,
"n_cpu_moe": 0,
"split_mode": "layer",
"tensor_parallel_size": 0,
"main_gpu": 0,
"no_kv_offload": false,
"flash_attn": 1,
"devices": "auto",
"tensor_split": "0.00",
"tensor_buft_overrides": "per_layer_token_embd=CPU",
"load_mode": "none",
"lazy_mode": "off",
"embeddings": false,
"no_op_offload": 0,
"no_host": false,
"no_repack": false,
"fit_target": 0,
"fit_min_ctx": 0,
"n_prompt": 0,
"n_gen": 512,
"n_depth": 0,
"test_time": "2026-08-30T20:28:13Z",
"avg_ns": 16168556612,
"stddev_ns": 0,
"avg_ts": 31.666401,
"stddev_ts": 0.000000,
"samples_ns": [ 16168556612 ],
"samples_ts": [ 31.6664 ]
}
]
~ $