Intel · Xe2 (Battlemage) · 2026
Intel Arc Pro B70 for local AI
32GB of new VRAM for well under NVIDIA or R9700 prices, held back by Intel's maturing software.
Updated · Speeds marked ~ are estimates; seemethodology
- VRAM
- 32GB
- Bandwidth
- 608 GB/s
- Power
- 230W
- Runs well
- 8 models
Arc Pro B70
Typically about $1,300
Can the Arc Pro B70 run LLMs?
The Arc Pro B70 has 32GB of GDDR6 ECC and 608 GB/s of memory bandwidth, which decides how big a model fits and how fast it generates text. It comfortably runs models up to about 35B parameters: the largest in our catalog is Qwen3.6 35B-A3B at Q5_K_M (28GB with 8K context, ~62 tokens/sec). Llama 3.3 70B Instruct only runs with partial CPU offload, at about 2.3 tokens/sec with CPU offload.
Recommendations
Best models for the Arc Pro B70
Best overall · also coding, reasoning, vision, fastest, largest that fits
Qwen3.6 35B-A3BQ5_K_M · 28GB · ~62 tokens/sec
Runs wellRunner-up
Qwen3.8 27BQ6_K · 25GB · ~14 tokens/sec
Runs wellStretch pick (CPU offload)
Qwen3.5 122B-A10BQ4_K_M · 77GB · about 10 tokens/sec with CPU offload
Partial offload
Compatibility
What runs on the Arc Pro B70
Every current model in our catalog, at the best quantization that fits in 32GB with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.
| Model | Parameters | Best quant | Memory | Speed | Verdict |
|---|---|---|---|---|---|
| Gemma 4 E4B | 8B | Q8_0 | 10 GB | ~33 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12 GB | ~29 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18 GB | ~21 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | ~59 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q6_K | 23 GB | ~48 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q6_K | 25 GB | ~14 tok/s | Runs well |
| Gemma 4 31B | 31B | Q5_K_M | 26 GB | ~14 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q5_K_M | 28 GB | ~62 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q4_K_M | 68 GB | ~8.8 tok/s | Partial offload |
| Llama 4 Scout | 109B (17B active) | Q4_K_M | 70 GB | ~6.3 tok/s | Partial offload |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68 GB | ~21 tok/s | Partial offload |
| Qwen3.5 122B-A10B | 122B (10B active) | Q4_K_M | 77 GB | ~10 tok/s | Partial offload |
| DeepSeek V4 Flash | 284B (13B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3-Flash | 320B (18B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3 | 744B (40B active) | — | — | — | Won't fitWhat it needs |
Performance
Arc Pro B70 vs. alternatives on Qwen3.8 27B
Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.
RTX 5090
Q6_K · 32GB
~59tok/s
R9700
Q6_K · 32GB
~18tok/s
Arc Pro B60
Q4_K_M · 24GB
~15tok/s
Arc Pro B70
Q6_K · 32GB
~14tok/s
Measured
Arc Pro B70 benchmark results
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| Llama 2 7B Chat | Q4_0 | llama.cpp (Vulkan) | 110 tok/s | 3380 tok/s | llama.cpp discussion #10879: Performance of llama.cpp with Vulkan · October 2026 |
Our take
Should you buy the Arc Pro B70 for AI?
The Arc Pro B70 is Intel's larger Battlemage card, with 32 Xe2 cores and 32GB of ECC GDDR6 on a 256-bit bus for 608 GB/s. It launched in March 2026 at $949, the lowest price of any new 32GB card at the time. Reviews found the hardware capable for vLLM and llama.cpp inference, but Intel's software stack still needs more setup and has gaps compared with CUDA and ROCm. Shortage pricing raised it to about $1,300 by October 2026, still below the R9700.
What we like
- 32GB with ECC for less than rival 32GB cards
- 608 GB/s bandwidth
- 230W board power and a PCIe 5.0 x16 link
Watch out for
- Intel's AI software stack is still maturing
- Price rose well above the $949 MSRP
Best for: Multi-GPU builds that need lots of VRAM per dollar · Running 32B models at 6 bit on one card
Specifications
Arc Pro B70 specs that matter for AI
- Memory
- 32GB GDDR6 ECC
- Memory bandwidth
- 608 GB/s
- Memory bus
- 256-bit
- Architecture
- Xe2 (Battlemage)
- Power
- 230W
- Interface
- PCIe 5.0 x16
- Released
- 2026
- Launch price
- $949
- Street price (new)
- ~$1,300 (2026-10)
- Software
- SYCL, Vulkan
Sources: StorageReview Arc Pro B70 review, Puget Systems Arc Pro B70 review, GPU Poet Arc Pro B70 specs and pricing
Cross-shop
Alternatives to the Arc Pro B70
Frequently asked questions
How much VRAM does the Arc Pro B70 have?
The Intel Arc Pro B70 has 32GB of GDDR6 ECC with 608 GB/s of memory bandwidth on a 256-bit bus.
What is the largest LLM the Arc Pro B70 can run?
Fully in GPU memory, the largest model in our catalog that runs comfortably is Qwen3.6 35B-A3B (35B (3B active)) at Q5_K_M, using about 28GB with 8K context at ~62 tokens/sec.
Can the Arc Pro B70 run Llama 3.3 70B?
Only with partial CPU offload. At Q4_K_M the model needs about 47GB with 8K context, so part of it spills into system RAM and speed drops to about 2.3 tokens/sec with CPU offload. For usable 70B speeds you want 48GB of VRAM or more (for example two 24GB cards).
What is the best LLM to run on the Arc Pro B70?
Our top pick is Qwen3.6 35B-A3B at Q5_K_M, which fits in 32GB with room for 8K of context and runs at ~62 tokens/sec. See our full list of the best models for the Arc Pro B70 for coding, reasoning, and vision picks.
Is the Arc Pro B70 good for local AI in 2026?
32GB of new VRAM for well under NVIDIA or R9700 prices, held back by Intel's maturing software.