Awesome Arc Pro B70

Arc Pro B70 leaderboard

Fastest submitted runs on the Intel Arc Pro B70, from localmaxxing.com (568 approved B70 runs, updated 2026-10-17 by a daily GitHub Action). Results are submitted by users and mostly not independently verified. Engines measure decode differently (llama-bench tg128, server inter-token medians, vLLM benchmarks), so compare within an engine. Best run per model, engine, quant and speculative-decoding setting.

Back to the hub · Prebuilt llama.cpp for the B70

How to reproduce the top runs

Single stream, one B70

# decode tok/s model quant engine spec ctx prefill tok/s date links
1 248.3 Qwen3.6-35B-A3B-GPTQ-Int4 GPTQ-4bit vllm vLLM 0.26.1.dev0 XPU, built from… MTP (cmd) 32768 12400 2026-08-08 run
2 244.5 NVIDIA-Nemotron-3.5-Lightning-30B-A3B-B… gptq-int4 vllm 0.26.1   2048 4511 2026-09-30 run
3 244.5 NVIDIA-Nemotron-3.5-Lightning-30B-A3B-W… gptq-int4 vllm 0.26.1   2048   2026-09-30 run
4 204.6 Qwen3.6-35B-A3B GPTQ-Int4 vllm 0.21.1.dev18+gec426505f XPU MTP (cmd) 16384 8715 2026-08-07 run
5 191.7 Qwen3.5-4B-quantized.w4a16 W4A16 vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-12 run
6 186.6 NVIDIA-Nemotron-3.5-Lightning-30B-A3B-B… GPTQ-INT4-G64-sym… vllm v0.26.1rc1.dev668+g3ee2df303 (XP… spec (cmd) 16384 7160 2026-08-13 run
7 132.4 LFM2.5-2.6B-GGUF Q8_0 llama.cpp 9fee29e9435f; oneAPI 2026.1…   8192   2026-08-26 run · repro
8 131.5 Ornith-1.5-35B-A3B-GGUF Q4_K_M llama.cpp 9fee29e9435f + neural.downl…   8192   2026-08-26 run · repro
9 125.0 gemma-4-26B-A4B-it-GGUF Q8_K_XL llama.cpp c926ad098 local B70 SYCL/AO… MTP 32768   2026-07-01 run
10 124.1 Qwen3.5-9B-quantized.w4a16 W4A16 vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-12 run
11 120.8 Qwen3-4B Q4_K_M llama.cpp   2048 3326 2026-08-16 run
12 115.2 gemma-4-26B-A4B-it-GGUF UD-Q8_K_XL llama.cpp 9769 (c926ad098), IntelLLVM… MTP 32768 997 2026-07-03 run · repro
13 112.7 Qwen3.8-27B GPTQ-Int4+draft-I… vllm 0.27.2rc1.dev77+gac7509e2b XPU MTP (cmd) 131072 1696 2026-08-19 run
14 109.9 Qwen3.8-27B-GPTQ-Int4-asym GPTQ-Int4 vllm 0.31.0 XPU + B70 cookbook patche… MTP 65536 1829 2026-10-16 run · repro
15 108.4 Ornith-1.5-35B-A3B-GPTQ-Int4-sym-G128-M… GPTQ-Int4 vllm   2048 9073 2026-08-21 run
16 108.1 Qwen3-Coder-30B-A3B-Instruct-GGUF GGUF UD-Q4_K_XL llama.cpp llama.cpp fdb1db877 / llama…   4096   2026-07-04 run
17 107.5 Qwen3-30B-A3B-Instruct-2507-GGUF GGUF UD-Q4_K_XL llama.cpp llama-server version 9763 (…   4096   2026-07-04 run
18 106.7 Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 GPTQ-Int4 vllm   131072 1694 2026-08-19 run
19 106.5 Nex-N2-mini Q4_K_M llama.cpp   2048   2026-06-20 run
20 104.3 gemma-4-26B-A4B-it-GGUF Q8_K_XL llama.cpp c926ad098 local B70 SYCL/AO… draft-mtp 8192   2026-06-27 run
21 100.5 Qwen3-30B-A3B-Instruct-2507 Q4_K_XL llama.cpp   2048 1464 2026-08-16 run
22 98.1 Qwen3.5-9B-FP8-dynamic FP8 vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-07 run
23 97.8 Qwen3.8-27B-Uncensored-int4-AutoRound GPTQ-Int4 vllm 0.27.2rc1.dev78+g8e6d8e4f6 XPU (… spec (cmd) 2048 1782 2026-08-20 run
24 96.5 microsoft_Phi-4-mini-instruct-GGUF GGUF Q4_K_M llama.cpp fdb1db877   4096   2026-07-04 run
25 95.1 Qwen3.8-27B exl3 vllm 0.26.1   2048   2026-09-30 run
26 94.6 QUEST-30B-RL-B70-Turbo-GGUF Q5_K_M llama.cpp   2048 1342 2026-08-17 run
27 93.0 NVIDIA-Nemotron-3.5-Lightning-30B-A3B-B… GPTQ-INT4-G64-sym… vllm v0.26.1rc1.dev668+g3ee2df303 (XP…   16384 8368 2026-08-12 run
28 87.8 Laguna-XS-2.1 Q4_K_M llama.cpp   2048 1044 2026-08-16 run
29 87.4 NVIDIA-Nemotron-3.5-Lightning-30B-A3B-G… Q4_0 llama.cpp b306 SYCL (f785fc9ea) + cus… MTP (cmd) 16384 766 2026-08-12 run
30 87.4 NVIDIA-Nemotron-3.5-Lightning-30B-A3B-B… Q4_0 llama.cpp b306 SYCL (f785fc9ea) + cus… MTP (cmd) 16384 766 2026-08-11 run

Single stream, multiple B70s

# decode tok/s model quant engine spec ctx prefill tok/s date links
1 236.9 Qwen3.5-4B-quantized.w4a16 W4A16 vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-08 run
2 172.3 Qwen3.5-9B-quantized.w4a16 W4A16 vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-07 run
3 161.9 Muse-Glimmer-30B-GGUF UD-Q8_K_XL llama.cpp 10433; 030ebb558 + Muse TP4… draft-dflash 32768   2026-08-14 run
4 147.8 Qwen3.5-9B-FP8-dynamic FP8 vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-07 run
5 125.5 Laguna-S-2.1-INT4 INT4_W4A16 vllm 0.1.dev1172+g4a6fd8747; local vL… dflash 8192   2026-08-01 run
6 122.2 Qwen3.8-Flash-Next-GSQ-RCO-GGUF IQ3_XXS strata 0.1.44.1 MTP 2048   2026-10-16 run
7 120.8 Qwen3.6-35B-A3B BF16 (native) vllm vllm-main d272418f4 (XPU graph +…   65536   2026-06-25 run
8 117.5 Qwen3.8-27B-int4-AutoRound int4 AutoRound W4… vllm vllm 0.27.2rc1.dev77+gac7509e2b … MTP 1024   2026-09-12 run
9 110.7 Qwen3.6-35B-A3B FP8 (online) + FP… vllm vllm-main d272418f4 (XPU graph +…   65536   2026-06-25 run
10 102.5 Qwen3.6-35B-A3B BF16 vllm g6607a80da (custom)   262144   2026-06-19 run

Batched / concurrent throughput (aggregate tok/s)

# decode tok/s model quant engine spec ctx prefill tok/s date links
1 1139.8 Qwen3.6-35B-A3B GPTQ-Int4 vllm 0.21.1.dev18+gec426505f XPU   16384 8715 2026-08-07 run
2 796.2 gemma-4-12B-it-int4-AutoRound INT4 AutoRound W4… vllm 0.20.2rc1.dev2+gc51df4300.d20260…   32768   2026-06-07 run
3 438.5 Qwen3.6-27B sym_int4 vllm   2048 10 2026-08-16 run
4 335.5 QUEST-30B-RL-B70-Turbo-GGUF Q5_K_M GGUF llama.cpp SYCL moe-ready build used b…   131072 76 2026-07-04 run
5 283.1 Laguna-XS-2.1-GGUF Q5_K_M GGUF llama.cpp SYCL moe-ready build used b…   131072 66 2026-07-05 run
6 224.9 Ornith-1.0-35B-B70-Turbo Q5_K_M GGUF llama.cpp SYCL moe-ready build used b…   131072 33 2026-07-04 run
7 224.3 Nex-N2-mini-B70-Turbo-GGUF Q5_K_M GGUF llama.cpp SYCL moe-ready build used b…   131072 33 2026-07-04 run
8 224.2 Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 GPTQ-Int4 vllm   2048 5 2026-08-19 run
9 223.7 SIQ-1-35B-B70-Turbo-GGUF Q5_K_M GGUF llama.cpp SYCL moe-ready build used b…   131072 33 2026-07-04 run
10 209.0 Qwen-AgentWorld-35B-A3B-B70-Turbo-GGUF Q5_K_M GGUF llama.cpp SYCL moe-ready build used b…   30720 1925 2026-07-04 run

Top single-card runs in detail

1. 248.3 tok/s: palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4 on vllm vLLM 0.26.1.dev0 XPU, built from…

vllm serve /models/Qwen3.6-35B-A3B-GPTQ-Int4 \
  --dtype float16 --max-model-len 32768 --max-num-seqs 16 \
  --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --block-size 64 \
  --language-model-only --trust-remote-code \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":4}'

2. 244.5 tok/s: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 on vllm 0.26.1

# No command snippet available

MoE vs Dense matched single-card comparison. p512/g128 greedy benchmark on Intel Arc Pro B70 at 200W cap. Yields 244.5 tok/s.

3. 244.5 tok/s: useful-quants/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-W4A16 on vllm 0.26.1

# No command snippet available

MoE vs Dense matched single-card comparison. p512/g128 greedy benchmark on Intel Arc Pro B70 at 200W cap. Yields 244.5 tok/s.

4. 204.6 tok/s: Qwen/Qwen3.6-35B-A3B on vllm 0.21.1.dev18+gec426505f XPU

vllm serve /model --quantization gptq --dtype float16 --max-model-len 16384 --max-num-seqs 64 --max-num-batched-tokens 8192 --served-model-name Qwen3.6-35B-A3B-MTP-Preserved-GPTQ-Int4 --language-model-only --speculative-config {"method":"mtp","num_speculative_tokens":4}

Single-stream MTP4 num_spec=4. vLLM 0.21 XPU int4moe + 2 patches: native int4 int8 store, BF16 MTP draft. Recurrent single-layer MTP emits 4 draft tokens/step. short/32 decode 204.6 t/s (grid, prefix-cache-off, warmup discarded). Prefill p4k 8715 t/s @230W. Decode @165W. –max-num-batched-tokens 8192. Model: Qwen3.6-35B-A3B-MTP-Preserved-GPTQ-Int4. 2026-08-07.

5. 191.7 tok/s: RedHatAI/Qwen3.5-4B-quantized.w4a16 on vllm vllm 0.27.2rc1.dev77+gac7509e2b …

MODEL_DIR=/path/to/Qwen3.5-4B-quantized.w4a16 VLLM_CACHE_DIR=/path/to/new-cache MTP_DEPTH=3 MAX_MODEL_LEN=1024 MAX_NUM_SEQS=1 repro/qwen35-4b-w4a16-b70/scripts/run-qwen35-4b-w4a16-server.sh  (ghcr.io/steveseguin/vllm-openai-xpu-qwen38-int4@sha256:78bd728d6109…)

Data: localmaxxing.com public API (/api/speed-tests?gpuName=Intel Arc Pro B70), user-submitted, provided as is. Generator: scripts/leaderboard.py.

This site is open source. Improve this page.