Local AI Registry

Recipe

qwen3-6-35b-a3b-bf16-native-intel-arc-pro-b70-32gb-vllm-tp4

qwen3-6-35b-a3b-bf16-native-intel-arc-pro-b70-32gb-vllm-tp4
JSON APIFind compatible recipes

Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.

Record

Status
candidate
Source
localmaxxing
Engine
vllm
Engine version
vllm-main d272418f4 (XPU graph + PatchB)
Graph
full-decode-only
Accelerators
4
Tensor parallel
4
Context tokens
65,536
Max concurrency
8
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/Qwen/Qwen3.6-35B-A3B
Repository
Qwen/Qwen3.6-35B-A3B
Status
known
Link type
Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Observed configuration

vllm

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmqszbv1q012hqr01fsmv2wnu

Observed source tokens

Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.

  1. vllm
  2. serve
  3. qwen36-35b-a3b-bf16-hf
  4. --tensor-parallel-size
  5. 4
  6. --dtype
  7. bfloat16
  8. --enable-prefix-caching
  9. --mamba-cache-mode
  10. align
  11. --max-model-len
  12. 65536
  13. --max-num-seqs
  14. 8
  15. --gpu-memory-utilization
  16. 0.90
  17. --max-num-batched-tokens
  18. 8192
  19. --trust-remote-code
  20. --enable-auto-tool-choice
  21. --tool-call-parser
  22. qwen3_coder
  23. --compilation-config
  24. {"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,4,8]}
FlagValue
--tensor-parallel-size4
--dtypebfloat16
--mamba-cache-modealign
--max-model-len65536
--max-num-seqs8
--gpu-memory-utilization0.90
--max-num-batched-tokens8192
--tool-call-parserqwen3_coder
--compilation-config{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,4,8]}

Environment

VariableValue
VLLM_XPU_ENABLE_XPU_GRAPH1
ZE_AFFINITY_MASK0,2,3,4

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
—65,536—120.8—observedqwen3-6-35b-a3b-bf16-native-intel-arc-pro-b70-32gb-vllm-tp4-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count
4
hardware id
intel-arc-pro-b70-32gb
id
qwen3-6-35b-a3b-bf16-native-intel-arc-pro-b70-32gb-vllm-tp4
model instance id
qwen-qwen3-6-35b-a3b--bf16-native
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-6-35b-a3b-bf16-native-intel-arc-pro-b70-32gb-vllm-tp4-sweep
status
candidate

capabilities

engine

graph mode
full-decode-only
name
vllm
version
vllm-main d272418f4 (XPU graph + PatchB)

serving

max concurrency
8
max context tokens
65,536
tensor parallel
4
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
engine.graph mode · provenance · captured at
2026-09-01T01:32:29Z
engine.graph mode · reason explicit-cudagraph-mode-in-observed-command
engine.graph mode · state known
serving.kv cache tokens · provenance · captured at
2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
serving.max concurrency · provenance · captured at
2026-09-01T10:20:39Z
serving.max concurrency · reason explicit-source-server-capacity
serving.max concurrency · state known
serving.max context tokens · provenance · captured at
2026-09-01T10:20:39Z
serving.max context tokens · reason explicit-source-context-limit
serving.max context tokens · state known

metadata

localmaxxing · backend
xpu
localmaxxing · batch size
1
localmaxxing · hardware label
Intel Arc Pro B70
localmaxxing · notes
Qwen3.6-35B-A3B NATIVE BF16 (no quantization) on 4x Intel Arc Pro B70 (TP=4) via vLLM-XPU graph-mode. b1 DECODE = 120.76 t/s (corrected after-first-token, mean-of-5, temp 0, deterministic 120.71-120.81). FULL_DECODE_ONLY XPU graph capture (4-way XCCL allreduce captured cleanly) on vllm-main d272418f4 + locally-patched cutlass work_group_scratch attention kernel. Cold prefill ~6500 tok/s (5.3k-tok prompt); align prefix-cache warm re-prefill 145ms vs 810ms cold (~5.6x). Higher-batch aggregate (single TP=4 instance): b8=593, b16=602, b24=596 tok/s. Full-precision BF16 BEATS our own online-FP8 TP=2 (110.73 b1) by ~9% on decode AND beats the prior B70 BF16-TP4 community best (102.5) by +18 t/s -- FP8 dequant overhead exceeds its bandwidth savings on this 3B-active MoE. qwen3_coder tool-calling verified. No power-limit changes; high-power B70 variant card excluded (4 matched regular-power cards).
localmaxxing · observed command
ZE_AFFINITY_MASK=0,2,3,4 VLLM_XPU_ENABLE_XPU_GRAPH=1 vllm serve qwen36-35b-a3b-bf16-hf --tensor-parallel-size 4 --dtype bfloat16 --enable-prefix-caching --mamba-cache-mode align --max-model-len 65536 --max-num-seqs 8 --gpu-memory-utilization 0.90 --max-num-batched-tokens 8192 --trust-remote-code --enable-auto-tool-choice --tool-call-parser qwen3_coder --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,4,8]}'
localmaxxing · run id
cmqszbv1q012hqr01fsmv2wnu
localmaxxing · tokenized · arguments
vllm, serve, qwen36-35b-a3b-bf16-hf, --tensor-parallel-size, 4, --dtype, bfloat16, --enable-prefix-caching, --mamba-cache-mode, align, --max-model-len, 65536, --max-num-seqs, 8, --gpu-memory-utilization, 0.90, --max-num-batched-tokens, 8192, --trust-remote-code, --enable-auto-tool-choice, --tool-call-parser, qwen3_coder, --compilation-config, {"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,4,8]}
localmaxxing · tokenized · environment · VLLM XPU ENABLE XPU GRAPH
1
localmaxxing · tokenized · environment · ZE AFFINITY MASK
0,2,3,4
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:10:02Z

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmqszbv1q012hqr01fsmv2wnu ↗