BBrainOutput
AMD·Datacenter acceleratorProvisional

AMD Instinct MI325X 256GB: Specs & Local-AI Compatibility

256GB HBM3e accelerator — very large memory for big-model serving.

Some details here are provisional (placeholder). Treat specs as approximate and verify against the manufacturer before relying on them or purchasing.

Specs

Memory
256 GB
Memory type
HBM3e
Bandwidth
6,000 GB/s
Approx FP16
to verify
Architecture
CDNA 3
Process
TSMC
Power
1,000 W
Launch
2024

Models this chip can run

Open models graded for a single AMD Instinct MI325X 256GB, best fit first.

  • Qwen3 235B-A22B (MoE)
    Qwen · ~235B · 128K ctx · Apache-2.0

    Fits at Q4_K_M (~130GB) with ~95.3GB headroom — about 1 concurrent instance.

    Q4_K_M · ~130GBRuns well
  • Qwen2.5 72B
    Qwen · ~72B · 128K ctx · Qwen License

    Fits at FP16 (~145GB) with ~80.3GB headroom — about 1 concurrent instance.

    FP16 · ~145GBRuns well
  • Llama 3.1 70B
    Llama · ~70B · 128K ctx · Llama Community License

    Fits at FP16 (~140GB) with ~85.3GB headroom — about 1 concurrent instance.

    FP16 · ~140GBRuns well
  • Llama 3.3 70B
    Llama · ~70B · 128K ctx · Llama Community License

    Fits at FP16 (~140GB) with ~85.3GB headroom — about 1 concurrent instance.

    FP16 · ~140GBRuns well
  • DeepSeek-R1 Distill Llama 70B
    DeepSeek · ~70B · 128K ctx · MIT

    Fits at FP16 (~140GB) with ~85.3GB headroom — about 1 concurrent instance.

    FP16 · ~140GBRuns well
  • Mixtral 8x7B (MoE)
    Mistral · ~47B · 32K ctx · Apache-2.0

    Fits at FP16 (~90GB) with ~135.3GB headroom — about 2 concurrent instances.

    FP16 · ~90GBRuns well
  • CodeLlama 34B
    CodeLlama · ~34B · 16K ctx · Llama Community License

    Fits at FP16 (~68GB) with ~157.3GB headroom — about 3 concurrent instances.

    FP16 · ~68GBRuns well
  • Qwen2.5 32B
    Qwen · ~32B · 128K ctx · Apache-2.0

    Fits at FP16 (~64GB) with ~161.3GB headroom — about 3 concurrent instances.

    FP16 · ~64GBRuns well

Build a private AI Business OS on AMD Instinct MI325X 256GB

Run your own AI agents on hardware you control — private by design, no per-seat data leaving your premises. BrainOutput helps you pick the right machine and turn it into a working AI Business OS.

Explore the AI Business OS