Llama 4 Maverick 402B · llama.cpp · UD-Q2_K_XL

status: blocked

Configuration

Modelmeta-llama/Llama-4-Maverick-17B-128E-Instruct
CompanyMeta
FamilyLlama
Parameters402B / 17B (MoE)
Enginellama.cpp
Quant / precisionUD-Q2_K_XL
Why this quant2-bit (UD-Q2_K_XL, ~122 GB) is the smallest published quant. It technically fits the 128 GB box but leaves essentially no room for the CUDA context + KV cache, so it would OOM at any real context length — hence blocked for a human fit decision rather than a guaranteed run.
Downloadunsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF
Context window131072
Input modalitiestext

Measured results

Prefill tok/s
Decode tok/s
Peak memory (GB)

Full run command

# blocked — ~122 GB weights leave ~no KV headroom on 128 GB; needs a fit decision

llama-4-maverick-17b-128e Meta Llama UD-Q2_K_XL 130B+ conc-32

Notes

Why this is interesting (and why it’s blocked): a 402B MoE (17B active) running on one tiny box is a great headline — at Unsloth’s dynamic 2-bit it’s ~122 GB, which just fits 128 GB. The catch is there’s almost nothing left for KV cache, so it only runs at a trivially small context and risks OOM. It’s right at the edge: worth a human call on whether to attempt it (smaller-than-2-bit quant, minimal context) or skip. Compare with the Qwen3.5-397B-A17B · UD-IQ1_M config, which fits the same class of model with real headroom at ~1.7-bit.