I host 30+ free APEX MoE quantizations as independent research. My only local hardware is an NVIDIA DGX Spark (122 GB unified memory), enough for ~30-50B-class MoEs, but bigger ones (200B+) require rented compute on H100/H200/Blackwell, typically $20-100 per quant. If APEX quants are useful to you, your support directly funds those bigger runs.
APEX (Adaptive Precision for EXpert Models) quantizations of Jackrong/Qwopus3.6-35B-A3B-v1, with the MTP (multi-token prediction) head bundled for in-the-box self-speculative decoding.
These GGUFs bundle the model's MTP (multi-token prediction) head alongside the trunk in a single file, courtesy of llama.cpp PR #22673. With a recent llama.cpp (>= commit 255582687) you can enable self-speculative decoding using just this one file — no separate draft model needed:
Each quant is ~2.5% larger than its non-MTP counterpart (one extra transformer-block worth of weights, no embedding duplication since MTP shares the trunk's embed_tokens).
MTP draft head precision
The bundled MTP head (blk.40.* including the nextn.* projection + norms) is
quantized to Q8_0 (near-lossless) on every tier except I-Nano. I-Nano keeps
the trunk-tier precision on the MTP block (Q3_K routed experts, Q4_K attention)
but pins blk.40.nextn.eh_proj to Q4_K — see the explainer below.
This keeps draft accuracy high (important for spec-decode acceptance rate) at a
modest ~1 GB cost per file vs. trunk-tier precision.
Why the MTP head doesn't use imatrix
llama-imatrix runs normal forward passes that only activate the trunk
(blk.0..blk.39). The MTP head only fires during --draft-mtp spec decoding,
so its tensors get no imatrix activation data. We work around this by
quantizing the MTP head with static K-quant / Q8_0 which doesn't require
imatrix.
(A patch to llama-imatrix that records MTP activations during collection
is in progress at mudler/llama.cpp#mtp-imatrix
— once upstream this will let us push the drafter to lower bit-widths cleanly.)
What is APEX?
APEX is a MoE-aware mixed-precision quantization strategy. Per-tensor-role gradient: routed experts compress hardest, shared experts kept high (always active), attention/Mamba uniform; 5+5 symmetric edge gradient across the 40 trunk layers + MTP layer 40 at edge precision. I-variants use diverse imatrix calibration (chat, code, reasoning, tool-calling, agentic traces, Wikipedia).
