The AI Bench
12 GB · 456 GB/s

What can the Intel Arc B580 run?

Cheap 12 GB card. Works for small models, with more setup.

Best models for each task

Our top three for this machine. Chat and image tasks assume an 8K-token conversation; coding, documents and agents assume 32K.

Chat & writing13 of 25 run

  1. Gemma 4 12BRuns greatQ4_K_M · needs 8.5 GB at 8K · 35–55tok/sFast
  2. MiniCPM5 2BRuns greatQ8_0 · needs 4.0 GB at 8K · 85–130tok/sFast
  3. Qwen3.5 4BRuns greatQ8_0 · needs 6.4 GB at 8K · 55–85tok/sFast

Coding11 of 26 run

  1. Qwen3.5 4BRuns greatQ8_0 · needs 7.2 GB at 32K · 45–75tok/sFast
  2. Gemma 4 12BRuns wellQ4_K_M · needs 9.2 GB at 32K · 35–55tok/sFast
  3. Ornith 1.5 9BRuns greatQ4_K_M · needs 7.9 GB at 32K · 40–60tok/sFast

Long documents9 of 20 run

  1. Qwen3.5 4BRuns greatQ8_0 · needs 7.2 GB at 32K · 45–75tok/sFast
  2. MiniCPM5 2BRuns greatQ8_0 · needs 5.1 GB at 32K · 65–100tok/sFast
  3. Gemma 4 12BRuns wellQ4_K_M · needs 9.2 GB at 32K · 35–55tok/sFast

Agents & tools11 of 26 run

  1. Qwen3.5 4BRuns greatQ8_0 · needs 7.2 GB at 32K · 45–75tok/sFast
  2. MiniCPM5 2BRuns greatQ8_0 · needs 5.1 GB at 32K · 65–100tok/sFast
  3. Ornith 1.5 9BRuns greatQ4_K_M · needs 7.9 GB at 32K · 40–60tok/sFast

Images in5 of 13 run

  1. Gemma 4 12BRuns greatQ4_K_M · needs 8.5 GB at 8K · 35–55tok/sFast
  2. Qwen3.5 4BRuns greatQ8_0 · needs 6.4 GB at 8K · 55–85tok/sFast
  3. MiMo-V2.6 9BRuns greatQ4_K_M · needs 8.0 GB at 8K · 45–70tok/sFast

Make images4 of 6 run

  1. Stable Diffusion XL Base 1.0Runs greatFP16 · needs 7.9 GB
  2. FLUX.2 [klein] 4BRuns wellFP8 · needs 9.3 GB
  3. Z-Image-TurboRuns great4-bit GGUF · needs 5.0 GB

Speak text7 of 7 run

  1. Qwen3-TTS 1.7BRuns greatBF16 · needs 5.5 GB
  2. Kokoro-82MRuns greatFP32 · needs 1.4 GB
  3. Chatterbox Multilingual V3Runs greatFull · needs 4.2 GB

Transcribe6 of 7 run

  1. Whisper large-v3-turbo (via faster-whisper)Runs greatFP16 · needs 2.6 GB
  2. Nemotron 3.5 ASR Streaming 0.6BRuns greatFP32 · needs 3.5 GB
  3. Qwen3-ASR 1.7BRuns greatBF16 · needs 5.7 GB

Biggest model it runs well

Gemma 4 12B (Q4_K_M, 7.1 GB file) at about 35–55tok/sFast.

Should you buy it?

Buy it if

You want the cheapest new 12 GB card and like tinkering.

Skip it if

Local AI is your main use — NVIDIA software support is far better.

Watch out for

  • Intel archived its IPEX-LLM project in January 2026; use llama.cpp or LM Studio (Vulkan).

Measured on this hardware

Real runs from public sources, not estimates. All measurements.

ModelVersionPromptSpeedSource
Gemma 4 12BQ4_K_M51241.34 tok/ssource 2026-06-13