DeepSeek V3.2

685B parameters (Mixture-of-Experts, 37B active per token); an IQ2_XXS quant is about 217 GB, while a Q4_K_M or MXFP4 quant is about 341 to 352 GB.

• BARE MINIMUM

  Mac Studio M2 Ultra (192GB Unified Memory)

  192 GB UNIFIED MEMORY

  ~$3800 USED

  IQ2_XXS quant. Since the model requires 217 GB, it will swap heavily to SSD, resulting in a glacial speed of less than 1

  
• SWEET SPOT

  Mac Studio M3 Ultra (256GB Unified Memory)

  256 GB UNIFIED MEMORY

  ~$5500 STREET PRICE

  IQ2_XXS or Q2_K quant of about 217 GB fully in unified memory, delivering 5 to 10 tokens per second with a limited conte

  
• FULL POWER

  Mac Studio M3 Ultra (512GB Unified Memory)

  512 GB UNIFIED MEMORY

  ~$10000 USED

  Q4_K_M or MXFP4 quant of about 341 to 352 GB fully in unified memory, running at 15 to 20 tokens per second with up to 1

  

HARDWARE PICKS & PRICES ARE RESEARCHED FROM THE LIVE WEB AND REFRESHED AUTOMATICALLY. TREAT THEM AS BALLPARK, NOT GOSPEL.

STARS ARE THE COMMUNITY SCORE FOR THAT KIND OF WORK, DOCKED FOR HOW SQUASHED THE WEIGHTS GET AT EACH TIER. A DASH MEANS NO SPEED WAS RESEARCHED FOR THAT RIG. TIMES ARE COMPUTED FROM RESEARCHED SPEEDS. BALLPARK, NOT GOSPEL.

Users appreciate DeepSeek V3.2 for coding support and managing long documents, but some note it can be slow, make errors, or ignore system instructions in long chats.

655 POSTS · 343 COMMENTS · HACKER NEWS · GITHUB · X · REDDIT · DEV.TO · LEMMY · BLUESKY · DEV FORUMS · BLOG