Qwen3.8 2.4T A95B

2.4T parameters (95B active MoE), a 1-bit GGUF is about 397 GB

• BARE MINIMUM

  Mac Studio M3 Ultra (512GB Unified Memory)

  512 GB UNIFIED MEMORY

  ~$9,500 USED

  Too big for standard consumer hardware; runs 1-bit GGUF at 1 to 2 tok/s, extremely slow but fits entirely in memory

  
• SWEET SPOT

  8x RTX 3090 24GB GPUs (custom workstation)

  192 GB VRAM + 256 GB SYSTEM RAM

  ~$7,500 USED

  Dynamic 1-bit Standard (UD-IQ1_S) at 3 to 5 tok/s with partial GPU offloading

  
• FULL POWER

  8x NVIDIA H200 141GB GPUs (server node)

  1.1 TB VRAM

  ~$320,000 NEW

  FP8 or Q8_0 precision at 40 plus tok/s, full 262k context window

  

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 love Qwen3.8 2.4T A95B for its fast, accurate answers and strong coding assistance. However, many complain about high costs, frequent outages, and a tendency to lecture or refuse prompts.

688 POSTS · 316 COMMENTS · BLUESKY · REDDIT · HACKER NEWS · X · LEMMY · GITHUB · DEV.TO · BLOG · DEV FORUMS