Intel · Xe2 (Battlemage), BMG-G21 · 16 GB VRAM tier
Best AI Models for Intel Arc Pro B50 (16 GB)
Runs 7–9B models well; 14B fits but is slow (12 tok/s).
Low-power 16 GB SFF card (70 W) — modest 224 GB/s bandwidth caps throughput; best for compact, quiet inference rigs.
| # | Model | Speed | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 17.6 t/s | 17.6 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 34.4 t/s | 34.4 tok/s | Q4_K_M · 3.4 GB | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 35.8 t/s | 35.8 tok/s | Q4_K_M · 13.1 GB | |
| 4 | Gemma 2 9B ITGoogle Q4_K_M · 7.3 GB · 18.4 t/s | 18.4 tok/s | Q4_K_M · 7.3 GB | |
| 5 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 76.1 t/s | 76.1 tok/s | Q4_K_M · 5.5 GB |
| # | Model | Speed | Coding Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB · 35.8 t/s | 35.8 tok/s | Q4_K_M · 13.1 GB | |
| 2 | Q4_K_M · 7.9 GB · 21.1 t/s | 21.1 tok/s | Q4_K_M · 7.9 GB | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 17.6 t/s | 17.6 tok/s | overall | Q4_K_M · 6.6 GB |
| 4 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 76.1 t/s | 76.1 tok/s | Q4_K_M · 5.5 GB | |
| 5 | Gemma 4 E4B ITGoogle Q4_K_M · 5.6 GB · 21.1 t/s | 21.1 tok/s | Q4_K_M · 5.6 GB |
| # | Model | Speed | Reasoning Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 17.6 t/s | 17.6 tok/s | Q4_K_M · 6.6 GB | |
| 2 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 35.8 t/s | 35.8 tok/s | Q4_K_M · 13.1 GB | |
| 3 | Qwen3 8BAlibaba Q4_K_M · 7.6 GB · 20.3 t/s | 20.3 tok/s | Q4_K_M · 7.6 GB | |
| 4 | Q4_K_M · 7.9 GB · 21.1 t/s | 21.1 tok/s | Q4_K_M · 7.9 GB | |
| 5 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 76.1 t/s | 76.1 tok/s | Q4_K_M · 5.5 GB |
| # | Model | Speed | Long context Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 17.6 t/s | 17.6 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 34.4 t/s | 34.4 tok/s | Q4_K_M · 3.4 GB | |
| 3 | Q4_K_M · 7.9 GB · 21.1 t/s | 21.1 tok/s | Q4_K_M · 7.9 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 35.8 t/s | 35.8 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 76.1 t/s | 76.1 tok/s | Q4_K_M · 5.5 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB · 35.8 t/s | 35.8 tok/s | Q4_K_M · 13.1 GB | |
| 2 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 76.1 t/s | 76.1 tok/s | Q4_K_M · 5.5 GB | |
| 3 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 34.4 t/s | 34.4 tok/s | Q4_K_M · 3.4 GB | |
| 4 | Qwen3 8BAlibaba Q4_K_M · 7.6 GB · 20.3 t/s | 20.3 tok/s | Q4_K_M · 7.6 GB | |
| 5 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 17.6 t/s | 17.6 tok/s | Q4_K_M · 6.6 GB |
| # | Model | Speed | Vision Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 17.6 t/s | 17.6 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 34.4 t/s | 34.4 tok/s | Q4_K_M · 3.4 GB | |
| 3 | Gemma 3 4B ITGoogle Q4_K_M · 5.0 GB · 39.4 t/s | 39.4 tok/s | Q4_K_M · 5.0 GB | |
| 4 | Q4_K_M · 6.5 GB · 18.0 t/s | 18.0 tok/s | Q4_K_M · 6.5 GB | |
| 5 | MiniCPM V 4.7 1BOpenBMB Q4_K_M · 1.6 GB · 87.5 t/s | 87.5 tok/s | Q4_K_M · 1.6 GB |
The Arc Pro B50 runs 7–9B models well. Xe2 (Battlemage), BMG-G21 generation, released 2025.
Arc Pro B50 strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 37.8 tok/s in 15.7 GB
- Low power draw: 70 W board power; a 150 W PSU is enough
Weaknesses
- 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 16 GB here
- Slow on large dense models: Qwen3.8 27B at 7.5 tok/s
- Slow prompt reading: A 4K-token prompt on Qwen3.5 9B waits 5.2 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 (224 GB/s). Grey rows need more than 16 GB.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB17.6tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB11.8tok/s
- 7.5tok/s
- 37.8tok/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 1057 models on the Arc Pro B50
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 |
|---|---|---|---|---|---|---|
| Q4_K_M | 0.2 GB | 11200.0 t/s | — | S100, Runs great | ||
Opt 6.7BChat | FP16 | 14.7 GB | 7.6 t/s | 2K | D27, Barely runs | |
Granite 3.1 8B InstructChatFunctions | Q4_K_M | 7.9 GB | 20.2 t/s | 131K | B60, Decent | |
| Q4_K_M | 4.8 GB | 42.3 t/s | — | A76, Runs well | ||
Phi 3 Mini 128k InstructChatCode | Q4_K_M | 9.0 GB | 32.9 t/s | 131K | A72, Runs well | |
Darwin 36B OpusChatReasoningFunctions | Q2_K | 15.2 GB | 46.6 t/s | 262K | B55, Decent | |
Nemotron Mini 4B InstructChatFunctions | Q4_K_M | 3.2 GB | 37.7 t/s | 4K | A74, Runs well | |
Gemma 4 31BVision | IQ3_XS | 15.8 GB | 7.7 t/s | 262K | D15, Barely runs | |
| IQ2_XXS | 13.8 GB | 43.2 t/s | — | A69, Runs well | ||
Yi 6BChat | Q4_K_M | 4.2 GB | 27.5 t/s | 4K | A68, Runs well | |
Qwen3.6 35B A3B Claude 4.7 Opus Reasoning DistilledChatReasoning | Q2_K | 15.7 GB | 37.8 t/s | 262K | C44, Usable | |
SmolLM 135MChat | Q4_K_M | 0.4 GB | 260.5 t/s | 2K | S100, Runs great | |
| Q4_K_M | 5.8 GB | 22.4 t/s | 33K | B62, Decent | ||
| Q3_K_M | 14.3 GB | 8.1 t/s | 262K | C33, Usable | ||
Phi 3 Medium 128k InstructChatCode | Q4_K_M | 12.0 GB | 12.3 t/s | 131K | B51, Decent |
Specs
- Architecture
- Xe2 (Battlemage), BMG-G21
- VRAM
- 16 GB GDDR6
- Memory bandwidth
- 224.0 GB/s
- FP16 (dense)
- 85.00 TFLOPS
- Tensor cores
- 128
- TDP
- 70 W
- Released
- 2025-09-03
- MSRP
- $349
Arc Pro B50 FAQ
What is the largest model the Arc Pro B50 can run?
Gemma 4 26B A4B IT (25.8B) is the largest at 4-bit or better: at Q4_K_S it uses 15.5 GB of 16 GB, and runs at about 34.4 tok/s. At 3-bit, a bigger model fits too: Gemma 2 27B IT (27.2B) at IQ3_M, 15.0 GB.
Can the Arc Pro B50 run Qwen3.5 9B?
Yes. Qwen3.5 9B fits at Q4_K_M, using 6.6 GB of 16 GB, and it is the best model that runs well on this card.
Can the Arc Pro B50 run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 16 GB on this card.
How many tokens per second does the Arc Pro B50 give?
About 17.6 tok/s on Qwen3.5 9B at Q4_K_M, with the first token after 5.2 s on a 4K-token prompt.
What power supply does the Arc Pro B50 need?
Plan a 150 W PSU (2× the 70 W TDP).
Is 16 GB enough for coding and agents?
Yes for coding: GPT OSS 20B, the best of the 2 coding models that fit, runs at about 36 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.