← All GPUs
Intel · Workstation · Xe2 (Battlemage, 32 Xe-cores)
Intel Arc Pro B70
New since spring 2026: 32 GB for $949; cooler varies by partner board. Intel's quick reference lists 160-290 W board power across partner designs.
VRAM
32 GB
GDDR6 ECC
Memory bandwidth
608 GB/s
FP16 compute
not published
Price
~$949
new, maker or suggested price
What fits at 4-bit
The 4-bit file plus 8,192 tokens of context and 1.5 GB of overhead, against the VRAM of 1 and 2 cards. Speeds are measured medians where someone has published a run at that card count; a blank speed means no run, not a slow one. Machine pages carry estimates for the gaps.
| Model | Needs | 1 card (32 GB) | 2 cards (64 GB) |
|---|---|---|---|
| Qwen3.5 4B4B · Q4_K_M | 4.3 GB | fits | fits |
| Llama 3.1 8B Instruct8B · Q4_K_M | 7.1 GB | fits | fits |
| Qwen3 8B8.2B · Q4_K_M | 7.3 GB | fits | fits |
| Gemma 4 E4B IT8B · Q4_K_M | 6.5 GB | fits | fits |
| Granite 4.2 8B8.8B · Q4_K_M | 7.9 GB | fits | fits |
| Qwen3.5 9B9B · Q4_K_M | 7.0 GB | fits | fits |
| Gemma 4 12B IT11.95B · Q4_K_M | 8.2 GB | fits | fits |
| Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M | 10.4 GB | fits | fits |
| Phi-414.7B · Q4_K_M | 11.3 GB | fits | fits |
| gpt-oss-20b21B, 3.6B active · MXFP4 (native) | 12.5 GB | fits | fits |
| Devstral Small 2 24B (2512)24B · Q4_K_M | 16.1 GB | 33 tok/smedian of 2 runs | fits |
| Qwen3.6 27B27B · Q4_K_M | 17.7 GB | fits | fits |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M | 17.4 GB | 41 tok/smedian of 3 runs | fits |
| Qwen3.8 27B27B · Q4_K_M | 18.2 GB | 25 tok/smedian of 3 runs | fits |
| Granite 4.2 30B29.3B · Q4_K_M | 20.3 GB | fits | fits |
| Gemma 4 31B IT30.7B · Q4_K_M | 18.9 GB | fits | fits |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M | 19.5 GB | fits | fits |
| Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M | 21.9 GB | fits | fits |
| Qwen3.6 35B-A3B35B, 3B active · Q4_K_M | 22.4 GB | fits | fits |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M | 25.3 GB | fits | fits |
| Llama 3.3 70B Instruct70B · Q4_K_M | 43.6 GB | too large | fits |
| Qwen3 Coder Next80B, 3B active · Q4_K_M | 46.9 GB | too large | fits |
| gpt-oss-120b117B, 5.1B active · MXFP4 (native) | 60.2 GB | too large | fits |
| Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M | 62.4 GB | too large | fits |
Measured speeds
One card
| Model | Quant | Backend | Generation | Prompt | Context | Source |
|---|---|---|---|---|---|---|
| Devstral Small 2 24B (2512) | Q4_0 | llama.cpp b10566 SYCL | 34 tok/s | 972 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Devstral Small 2 24B (2512) | Q4_K - Medium | llama.cpp b10566 SYCL | 32 tok/s | 1,044 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Devstral Small 2 24B (2512) | Q5_K - Medium | llama.cpp b10566 SYCL | 30 tok/s | 1,074 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Devstral Small 2 24B (2512) | Q6_K | llama.cpp b10566 SYCL | 27 tok/s | 1,199 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Devstral Small 2 24B (2512) | Q8_0 | llama.cpp b10566 SYCL | 19 tok/s | 1,189 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Gemma 4 26B-A4B IT | UD-Q4_K_M | llama.cpp b8759 (Vulkan) | 42 tok/s | 1,502 tok/s512-token prompt | 0 | owner report2026-04 |
| Gemma 4 26B-A4B IT | UD-Q4_K_M | llama.cpp (build not stated, Vulkan) | 41 tok/s | 1,654 tok/s512-token prompt | 0 | owner report2026-04 |
| Gemma 4 26B-A4B IT | MXFP4 MoE | llama.cpp (build not stated, SYCL) | 24 tok/s | 674 tok/s512-token prompt | 0 | owner report2026-04 |
| Gemma 4 31B IT | UD-Q5_K_XL | llama.cpp (build not stated, SYCL) | 8.0 tok/s | – | 0 | review2026-04 |
| Qwen3.8 27B | IQ3_XXS - 3.0625 bpw | llama.cpp b10644 SYCL,BLAS | 10 tok/s | 445 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q4_0 | llama.cpp b10644 SYCL,BLAS | 27 tok/s | 813 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q4_K - Medium | llama.cpp b10644 SYCL,BLAS | 25 tok/s | 867 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q4_K - Medium | llama.cpp b10553 SYCL | 21 tok/s | 789 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (SYCL, build not stated) | – | 1,150 tok/s4,096-token prompt | 0 | llama.cpp benchmark discussion2026-10 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (SYCL, build not stated) | – | 870 tok/s64,121-token prompt | 0 | llama.cpp benchmark discussion2026-10 |
| Qwen3.8 27B | Q5_K - Medium | llama.cpp b10644 SYCL,BLAS | 22 tok/s | 1,031 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q5_K - Medium | llama.cpp b10553 SYCL | 21 tok/s | 950 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q6_K | llama.cpp b10644 SYCL,BLAS | 20 tok/s | 985 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q6_K | llama.cpp b10553 SYCL | 20 tok/s | 951 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q8_0 | llama.cpp (commit 63b64a50a, SYCL, GGML_SYCL_F16=ON) | 16 tok/s | 1,447 tok/s2,048-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q8_0 | llama.cpp (commit 63b64a50a, SYCL, GGML_SYCL_F16=ON) | – | 1,388 tok/s8,192-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q8_0 | llama.cpp (commit 63b64a50a, SYCL, GGML_SYCL_F16=ON) | – | 795 tok/s131,072-token prompt | 0 | llama.cpp benchmark discussion2026-08 |
Speculative decoding, several users, or long context
Kept apart because each measures a different workload from one person chatting near the start of a conversation.
| Model | Quant | Backend | Generation | Prompt | Context | Measured on | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (SYCL, build not stated) | – | 856 tok/s4,096-token prompt | 0 | 2× Arc Pro B70 | llama.cpp benchmark discussion2026-10 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (SYCL, build not stated) | – | 668 tok/s64,121-token prompt | 0 | 2× Arc Pro B70 | llama.cpp benchmark discussion2026-10 |
| Qwen3.8 27B | Q8_0 | llama.cpp (commit 63b64a50a, SYCL, GGML_SYCL_F16=ON) | 51 tok/s | – | 0 | Arc Pro B70 | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q8_0 | llama.cpp (commit 63b64a50a, SYCL, GGML_SYCL_F16=ON) | 34 tok/s | – | 0 | Arc Pro B70 | llama.cpp benchmark discussion2026-08 |
Specifications
- Memory
- 32 GB GDDR6 ECC
- FP16 basis
- Intel publishes 367 INT8 TOPS (XMX, dense) and 22.94 TFLOPS FP32; no FP16 figure.
- Board power
- 230 W
- PCIe
- Gen 5 x16
- NVLink
- No
- Software
- oneAPI/SYCL, Vulkan
- Released
- 2026
- Availability
- sold new
- Price basis
- Intel suggested starting price at the April 2026 launch.
- Maker page
- intel.com
Sources
- intel.com: 32 GB GDDR6 256-bit, 608 GB/s, 230 W TBP, PCIe 5.0 x16, 367 INT8 TOPS, ECC
- intel.com: Intel Arc Pro B-series quick reference (Mar 2026): B70/B65/B60/B50 memory, bandwidth, INT8 TOPS, board power
- phoronix.com: $949 suggested price, 230 W, available at launch as Intel-branded and partner cards