Intel · Xe2 (Battlemage), BMG-G31
Best AI Models for Intel Arc Pro B65 (32 GB)
Runs 27–31B and 35B MoE models well; 70B doesn't fit in 32 GB.
B60-class compute with B70 memory — 32 GB at 608 GB/s for about $910, so capacity per dollar is its whole argument. Add-in-board partners only; Intel ships no reference card and set no MSRP.
| # | Model | Speed | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 66.5 t/s | 66.5 tok/s | Q4_K_M · 16.5 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 75.1 t/s | 75.1 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Q4_K_M · 19.3 GB · 66.0 t/s | 66.0 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Coding Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 57.0 t/s | 57.0 tok/s | Q4_K_M · 22.0 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 66.5 t/s | 66.5 tok/s | Q4_K_M · 16.5 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 75.1 t/s | 75.1 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Q4_K_M · 19.3 GB · 66.0 t/s | 66.0 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Reasoning Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.5 27BAlibaba Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 66.5 t/s | 66.5 tok/s | Q4_K_M · 16.5 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 75.1 t/s | 75.1 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Q4_K_M · 19.3 GB · 66.0 t/s | 66.0 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Math Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 57.0 t/s | 57.0 tok/s | Q4_K_M · 22.0 GB | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 66.5 t/s | 66.5 tok/s | overall | Q4_K_M · 16.5 GB |
| 5 | Q4_K_M · 19.3 GB · 66.0 t/s | 66.0 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 93.3 t/s | 93.3 tok/s | Q4_K_M · 3.4 GB | |
| 2 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 75.1 t/s | 75.1 tok/s | Q4_K_M · 13.1 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 66.5 t/s | 66.5 tok/s | Q4_K_M · 16.5 GB | |
| 4 | Q4_K_M · 19.3 GB · 66.0 t/s | 66.0 tok/s | Q4_K_M · 19.3 GB | |
| 5 | Qwen3 30B A3B Thinking 2507Alibaba Q4_K_M · 19.4 GB · 61.1 t/s | 61.1 tok/s | Q4_K_M · 19.4 GB |
| # | Model | Speed | Vision Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 17.5 t/s | 17.5 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 66.5 t/s | 66.5 tok/s | Q4_K_M · 16.5 GB | |
| 4 | Gemma 3 27B ITGoogle Q4_K_M · 21.7 GB · 16.8 t/s | 16.8 tok/s | Q4_K_M · 21.7 GB | |
| 5 | MiniCPM V 4.7 35B A3BOpenBMB Q4_K_M · 22.1 GB · 62.3 t/s | 62.3 tok/s | Q4_K_M · 22.1 GB |
The Arc Pro B65 runs 27–31B and 35B MoE models well. Xe2 (Battlemage), BMG-G31 generation, released 2026.
Arc Pro B65 strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 57.0 tok/s in 22 GB
- Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 14.1 GB free
Weaknesses
- 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 32 GB here
- Slow prompt reading: A 4K-token prompt on Qwen3.8 27B waits 12.8 s for the first token
- oneAPI, not CUDA: No FP8 or NVFP4 weights and no vLLM AWQ or GPTQ; use GGUF builds
Output speed depends on memory bandwidth (608 GB/s). Grey rows need more than 32 GB.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB47.8tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB32.0tok/s
- 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB17.5tok/s
- 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB57.0tok/s
- 70BLlama 3.3 70B Instruct · not at 4-bitneeds 51.3 GB
- 120B MoEGPT OSS 120B · not at 4-bitneeds 70.8 GB
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
All 1176 models on the Arc Pro B65
Featured onlyBest quant that fits, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
FastVLM 7BChat | BF16 | 16.8 GB | 19.1 t/s | 33K | B59, Decent | |
| Q4_K_M | 7.4 GB | 55.4 t/s | 131K | A82, Runs well | ||
Gemma 3 270M ITChat | Q4_K_M | 1.2 GB | 1688.9 t/s | — | S100, Runs great | |
Qwen3 VL 2B ThinkingVisionFunctions | Q4_K_M | 3.5 GB | 168.0 t/s | 262K | S98, Runs great | |
Tiny NemotronHForCausalLM SuperChatFunctions | Q4_K_M | 0.3 GB | 30400.0 t/s | 262K | S100, Runs great | |
| Q4_K_M | 6.2 GB | 56.4 t/s | 8K | A82, Runs well | ||
Tiny NemotronHForCausalLM UltraChatFunctions | Q4_K_M | 0.3 GB | 30400.0 t/s | 262K | S100, Runs great | |
Pythia 160MChat | Q4_K_M | 0.1 GB | 2171.4 t/s | 2K | S100, Runs great | |
Saiga Llama3 8BChat | Q4_K_M | 6.2 GB | 56.4 t/s | 8K | A82, Runs well | |
Datarus R1 14B PreviewChatReasoning | Q4_K_M | 12.4 GB | 31.8 t/s | 131K | A71, Runs well | |
KONI Llama3.1 8B Instruct 20241024ChatFunctions | Q4_K_M | 7.3 GB | 56.4 t/s | 131K | A82, Runs well | |
Qwen3.5 2B BaseVision | Q4_K_M | 1.9 GB | 171.8 t/s | 262K | S98, Runs great | |
Granite 4.0 H TinyChatFunctions | Q4_K_M | 5.8 GB | 89.2 t/s | 131K | S89, Runs great | |
Nemotron Labs Diffusion 3BChatFunctions | Q4_K_M | 3.9 GB | 110.1 t/s | 262K | S93, Runs great | |
Typhoon Ocr 3BVision | Q4_K_M | 3.2 GB | 115.6 t/s | 128K | S94, Runs great |
Specs
- Architecture
- Xe2 (Battlemage), BMG-G31
- VRAM
- 32 GB GDDR6
- Memory bandwidth
- 608.0 GB/s
- FP16 (dense)
- 98.50 TFLOPS
- Tensor cores
- 160
- TDP
- 200 W
- Released
- 2026-03-25
- MSRP
- $910
Arc Pro B65 FAQ
What is the largest model the Arc Pro B65 can run?
Qwen3.6 35B A3B (36.0B) is the largest at 4-bit or better: at Q4_K_M it uses 22.0 GB of 32 GB, and runs at about 57.0 tok/s.
Can the Arc Pro B65 run Qwen3.8 27B?
Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of 32 GB, and it is the best model that runs well on this card.
Can the Arc Pro B65 run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 32 GB on this card.
How many tokens per second does the Arc Pro B65 give?
About 17.5 tok/s on Qwen3.8 27B at Q4_K_M, with the first token after 12.8 s on a 4K-token prompt.
What power supply does the Arc Pro B65 need?
Plan a 400 W PSU (2× the 200 W TDP).
Is 32 GB enough for coding and agents?
Yes for coding: Qwen3.8 27B, the best of the 9 coding models that fit, runs at about 18 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.