Qwen3.8 Flash Next vs Qwen3.8 27B

Web design bench, human-rated page prompts. One run each (latest complete run), 1 iteration. Pages are copied verbatim from the model output.

Qwen3.8 27B

Bench run
#40 (2026-09-15)
Model
Qwen3.8-27B
Quant
None, BF16 weights
KV cache
auto (BF16)
Backend
vLLM, TP4 (4 GPUs), MTP speculative decoding (2 tokens)
Context
262,144
Thinking
enabled, reasoning_effort xhigh
Sampling (sent)
temperature 0.7 (set by bench)
Server defaults
top_p 0.95, top_k 20, min_p 0, presence 0, repetition 1.0
Max tokens
250,000
Server args 
--trust-remote-code --dtype auto --kv-cache-dtype auto --reasoning-parser qwen3 --tensor-parallel-size 4 --enable-prefix-caching --performance-mode interactivity --max-num-seqs 4 --max-num-batched-tokens 24576 --enable-auto-tool-choice --tool-call-parser qwen3_coder --generation-config auto --override-generation-config {"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":0.0,"repetition_penalty":1.0} --speculative-config {"method":"mtp","num_speculative_tokens":2} --default-chat-template-kwargs {"enable_thinking":true,"preserve_thinking":true,"reasoning_effort":"xhigh"} --chat-template /home/shady/models/Qwen3.8-27B/chat_template.jinja --attention-backend FLASHINFER --gdn-prefill-backend flashinfer
gpu_memory_utilization 0.93, max_model_len 262144

Qwen3.8 Flash Next

Bench run
#42 (2026-09-17)
Model
Qwen3.8-Flash-Next-W4A16-AutoRound
Quant
Intel AutoRound W4A16 (calibrated; attention, linear attention and shared expert kept at BF16)
KV cache
auto (BF16, full precision)
Backend
vLLM, TP4 + expert parallel (4 GPUs)
Context
131,072
Thinking
enabled, reasoning_effort xhigh
Sampling (sent)
server defaults: temperature 1.0
Server defaults
top_p 0.95, top_k 20, min_p 0, presence 0, repetition 1.0
Max tokens
128,000
Server args 
--tensor-parallel-size 4 --enable-expert-parallel --disable-custom-all-reduce --max-num-seqs 2 --max-num-batched-tokens 2048 --kv-cache-dtype auto --compilation-config {"mode":"VLLM_COMPILE","cudagraph_mode":"FULL"} --no-enable-flashinfer-autotune --enable-prefix-caching --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_xml --default-chat-template-kwargs {"enable_thinking":true,"preserve_thinking":true,"reasoning_effort":"xhigh"} --override-generation-config {"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":0.0,"repetition_penalty":1.0} --enable-prompt-tokens-details --enable-per-request-metrics
gpu_memory_utilization 0.97, max_model_len 131072
env: VLLM_PLE_CPU_OFFLOAD=1, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, NCCL_P2P_DISABLE=0, NCCL_P2P_LEVEL=SYS
Prompt 27B 27B auto score 27B time / tokens Flash Next Flash Next auto score Flash Next time / tokens
Boutique hotel homepageOpen1.0015.3 min / 48,937Open0.929.5 min / 36,362
Conference landing pageOpen0.8316.0 min / 52,301Open0.8913.8 min / 51,513
Editorial magazine homepageOpen1.009.2 min / 29,394Open1.007.8 min / 29,093
SaaS landing heroOpen1.008.1 min / 25,814Open1.002.1 min / 7,910
Nonprofit fundraising campaign pageOpen1.0013.2 min / 41,997Open1.009.9 min / 36,728
Portfolio case-study pageOpen0.9017.2 min / 55,657Open1.007.6 min / 28,383
Pricing page with 3 tiersOpen1.007.7 min / 25,032Open1.000.5 min / 2,053
Open-ended product detail pageOpen1.0014.7 min / 44,281Open1.0013.5 min / 50,992
Real estate listing detail pageOpen0.9221.5 min / 66,191Open0.9213.0 min / 49,331
Restaurant reservation homepageOpen0.9210.8 min / 32,053Open0.5810.4 min / 38,966