Arc Pro B70 Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
The Arc Pro B70 is Intel's top 2026 workstation card with 32GB, running 30B-class MoE at about 113 tokens/s. Intel's vLLM is immature; use llama.cpp.
eBay
Arc Pro B70 price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full Arc Pro B70 specifications
- Memory32 GB · GDDR6
- Bandwidth608 GB/s
- FP16 / BF16183.5 TFLOPS
- INT8367 TOPS
- Bus width
- 256-bit
- Architecture
- Xe2 Battlemage (BMG-G31)
- Cores
- 4,096
- Tensor cores
- 256
- Ecosystem
- oneAPI
- Interface
- PCIe 5.0 x16
- Power
- 230 W
- Slots
- 2
- Type
- Workstation
- Launch date
- Mar 25, 2026
- MSRP
- $949
Best Arc Pro B70 deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedASRock B70 Intel Arc Pro B70 B70 CT 32GB 256-bit GDDR6 5.0 Creator New No Box$1,398Buy on eBay
UsedASRock Intel Arc Pro B70 Creator 32GB GDDR6 Graphics Card B70 CT 32G Open Box$1,399+ shippingBuy on eBay
NewASRock B70CT32G Arc Pro B70 Creator 32GB Workstation Graphics Card$1,559Buy on eBay
UsedIntel Arc Pro B70 32GB GDDR6 Intel Reference Graphics Card with Original Box$1,700$2,000-15%+$30 shippingBuy on eBay
NewInte Arc Pro B70 32GB 256-bit GDDR6 PCI Express 5.0 x16, 32 Ray Tracing Units,$1,850Buy on eBay
NewASRock B70 Intel Arc Pro B70 B70 CT 32GB 256-bit GDDR6 PCI Express 5.0 x16Top Rated$1,999Buy on eBay
What LLMs can a single Arc Pro B70 run?
A single B70 fits every model up to 35B at 4-bit, 30B-class MoE at about 110 tokens/s. At 8-bit only 27B dense fits, at 15 or 16.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 102.9 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 95.4 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 83.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 117.7 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 36.7 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 43.5 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 129.6 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 35.3 | 246K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 141.3 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 81.4 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 71.9 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 58.4 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 94.7 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 26.3 | 245K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 26.8 | 250K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 110 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 23.7 | 168K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 113.2 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 52.2 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 45.3 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 34.3 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 72.5 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 15.8 | 69K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 15.6 | 63K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 83.4 | 26K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 33 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can multiple Arc Pro B70 cards run?
Four B70s (128GB) are the usual setup: Qwen3.8 Flash Next at 4-bit, about 73 tokens/s. Two hold 70B-class but stay under 10. No NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 4 cards | $7,400 | 128 GB | 88.4 | 920 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | $7,400 | 128 GB | 69.5 | 920 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | $7,400 | 128 GB | 30.9 | 920 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | $7,400 | 128 GB | 54.9 | 920 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 4 cards | $7,400 | 128 GB | 73.1 | 920 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 8 cards | $14,800 | 256 GB | 49.7 | 1,840 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | $14,800 | 256 GB | 22.8 | 1,840 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | $14,800 | 256 GB | 34.3 | 1,840 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 8 cards | $14,800 | 256 GB | 48.1 | 1,840 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | $14,800 | 256 GB | 51.8 | 1,840 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- You want a 27B on one card at higher precision
Qwen3.8 27B at 8-bit (29 GB) fits, at about 16 tokens/s
- You mainly run MoE models
Qwen3.6 35B A3B and GLM 4.7 Flash run at about 110 tokens/s at 4-bit, and the 35B A3B can use the full 256K
- You're building a multi-GPU rig
Intel's own version is a dual-slot blower card with one 8-pin connector, and four of them give you 128GB
- You need ECC VRAM
Intel's spec page lists it as supported
Not a good fit
- Single-card generation speed matters most
On the same dense model it reaches less than 40% of an RTX 5090's speed
- You need FP8
This generation's matrix units don't support FP8
- You use only Windows but want vLLM
vLLM supports only Linux, and Intel's packaged version also requires Ubuntu 24.04
Before you buy
Intel's own version and board-partner versions
If you want to pack several cards closely, Intel's own version is the first choice: a dual-slot blower, 267 mm long, one 8-pin connector, 230W for the whole card. Board partners such as ASRock, Gunnir, Maxsun and Sparkle also make versions, with power settable between 160 and 290W; cooling and connectors depend on the specific model.
vLLM is Linux-only
For a single user, llama.cpp and Ollama are less hassle. vLLM requires Linux: Intel's packaged llm-scaler needs Ubuntu 24.04 and supports formats like AWQ, GPTQ and FP8, but not GGUF. In its July 2026 review, StorageReview called it beta quality at best: model support is limited, and some quantization methods don't work.
Try both llama.cpp backends
On Intel cards, llama.cpp has two backends, SYCL and Vulkan, and neither is consistently faster, so it's best to try both. In an April 2026 third-party test, Gemma 4 31B generated twice as fast on SYCL, but Qwen3.5 27B broke on SYCL (its second-turn answer got mixed up with the first-turn question) and only worked on Vulkan. OpenVINO 2026.1 has also added B70 support, but the backend it contributes to llama.cpp is still a preview and has only been validated on small models under 7B.
Compared with similar GPUs
Arc Pro B70 vs Radeon AI PRO R9700, RTX PRO 4500, and GeForce RTX 4080 SUPER 32GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | Arc Pro B70 | Radeon AI PRO R9700 |  RTX PRO 4500 | GeForce RTX 4080 SUPER 32GB |
|---|---|---|---|---|
| Memory | 32 GB | 32 GBSame | 32 GBSame | 32 GBSame |
| Bandwidth | 608 GB/s | 640 GB/s5% higher | 896 GB/s47% higher | 736 GB/s21% higher |
| BF16 / FP16 compute | 184 TFLOPS | 191 TFLOPS4% higher | 211 TFLOPS15% higher | 209 TFLOPS14% higher |
| INT8 compute | 367 TOPS | 383 TOPS4% higher | 422 TOPS15% higher | 418 TOPS14% higher |
| Qwen3.8 27B speed (2-bit) | 43.5 t/s | 45.6 t/s5% faster | 62.3 t/s43% faster | 52 t/s20% faster |
| Largest dense model (4-bit) | 51B | 51BSame | 51BSame | 51BSame |
| Power | 230 W | 300 W30% higher | 200 W13% lower | 320 W39% higher |
| Interface | PCIe 5.0 x16 | PCIe 5.0 x16 | PCIe 5.0 x16 | PCIe 4.0 x16 |
| Released | 2026-03 | 2025-07 | 2025-03 | 2024-01 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu