Intel · Xe2 (Battlemage) · 2024
Intel Arc B580 for local AI
The cheapest new 12GB card with good bandwidth, if you accept Intel's less mature software.
Updated · Speeds marked ~ are estimates; seemethodology
- VRAM
- 12GB
- Bandwidth
- 456 GB/s
- Power
- 190W
- Runs well
- 3 models
Arc B580
Typically about $330
Can the Arc B580 run LLMs?
The Arc B580 has 12GB of GDDR6 and 456 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 15B parameters: the largest in our catalog is Phi-4 at Q3_K_M (9.8GB with 8K context, ~32 tokens/sec). Our top all-round pick for it is Qwen3.5 9B (Q6_K, ~28 tokens/sec). Llama 3.3 70B Instruct only runs with partial CPU offload, at about 1.2 tokens/sec with CPU offload.
Recommendations
Best models for the Arc B580
Best overall · also coding, reasoning, vision, fastest
Qwen3.5 9BQ6_K · 10GB · ~28 tokens/sec
Runs wellLargest model that fits
Phi-4Q3_K_M · 10GB · ~32 tokens/sec
Runs wellStretch pick (CPU offload)
Qwen3.6 35B-A3BQ4_K_M · 24GB · about 29 tokens/sec with CPU offload
Partial offload
Compatibility
What runs on the Arc B580
Every current model in our catalog, at the best quantization that fits in 12GB 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 | ~25 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q6_K | 9.5 GB | ~28 tok/s | Runs well |
| Phi-4 | 14.7B | Q3_K_M | 9.8 GB | ~32 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | ~33 tok/s | Partial offload |
| Gemma 4 26B A4B | 26B (3.8B active) | Q4_K_M | 17 GB | ~30 tok/s | Partial offload |
| Qwen3.8 27B | 27B | Q4_K_M | 18 GB | ~5.3 tok/s | Partial offload |
| Gemma 4 31B | 31B | Q4_K_M | 22 GB | ~3.6 tok/s | Partial offload |
| Qwen3.6 35B-A3B | 35B (3B active) | Q4_K_M | 24 GB | ~29 tok/s | Partial offload |
| GLM-4.5-Air | 106B (12B active) | — | — | — | Won't fitWhat it needs |
| Llama 4 Scout | 109B (17B active) | — | — | — | Won't fitWhat it needs |
| gpt-oss-120b | 117B (5.1B active) | — | — | — | Won't fitWhat it needs |
| Qwen3.5 122B-A10B | 122B (10B active) | — | — | — | Won't fitWhat it needs |
| 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 B580 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.
Measured
Arc B580 benchmark results
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| Llama 2 7B Chat | Q4_0 | llama.cpp (Vulkan) | 70 tok/s | 620 tok/s | llama.cpp discussion #10879: Performance of llama.cpp with Vulkan · October 2026 |
Our take
Should you buy the Arc B580 for AI?
The Arc B580 pairs 12GB of GDDR6 with a 192-bit bus for 456 GB/s, more bandwidth than an RTX 3060 12GB or RTX 4060 Ti. It runs 8B models quickly and 14B models at 4 bit through llama.cpp's SYCL and Vulkan backends or Intel's OpenVINO. Software support is narrower than CUDA, and some tools need extra setup. New cards sold for about $330 in October 2026, up from a $249 launch MSRP.
What we like
- 12GB at 456 GB/s for a low new price
- 190W board power
- Works with llama.cpp SYCL and Vulkan backends
Watch out for
- Software support is less mature than CUDA
- 12GB limits model size
- PCIe 4.0 x8 link
Best for: Budget builds that want a new card · Running 8B to 14B models
Specifications
Arc B580 specs that matter for AI
- Memory
- 12GB GDDR6
- Memory bandwidth
- 456 GB/s
- Memory bus
- 192-bit
- Architecture
- Xe2 (Battlemage)
- Power
- 190W
- Interface
- PCIe 4.0 x8
- Released
- 2024
- Launch price
- $249
- Street price (new)
- ~$330 (2026-10)
- Street price (used)
- ~$300 (2026-10)
- Software
- SYCL, Vulkan
Sources: Intel Arc B580 product specifications, GPU Poet Arc B580 specs and pricing
Cross-shop
Alternatives to the Arc B580
Frequently asked questions
How much VRAM does the Arc B580 have?
The Intel Arc B580 has 12GB of GDDR6 with 456 GB/s of memory bandwidth on a 192-bit bus.
What is the largest LLM the Arc B580 can run?
Fully in GPU memory, the largest model in our catalog that runs comfortably is Phi-4 (14.7B) at Q3_K_M, using about 10GB with 8K context at ~32 tokens/sec.
Can the Arc B580 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 1.2 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 B580?
Our top pick is Qwen3.5 9B at Q6_K, which fits in 12GB with room for 8K of context and runs at ~28 tokens/sec. See our full list of the best models for the Arc B580 for coding, reasoning, and vision picks.
Is the Arc B580 good for local AI in 2026?
The cheapest new 12GB card with good bandwidth, if you accept Intel's less mature software.