license: gemma
base_model:

• google/gemma-4-26B-A4B-it-qat-q4_0-unquantized
• huihui-ai/Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated
• google/gemma-4-26B-A4B-it-qat-q4_0-unquantized-assistant
  base_model_relation: quantized
  language:
• ja
• en
  tags:
• gemma4
• moe
• nvfp4
• w4a4
• qat
• quantized
• abliterated
• vllm
• compressed-tensors
• blackwell
• sm120
• speculative-decoding
• mtp
• gemma4-assistant
  library_name: transformers
  pipeline_tag: image-text-to-text
  model_type: gemma4
  quantized_by: Lna-Lab

Huihui-gemma-4-26B-A4B-it-qat-abliterated-MTP-NVFP4

One repo, speculative decoding included. This is the MTP bundle of Huihui-gemma-4-26B-A4B-it-qat-abliterated-pad768-NVFP4: the NVFP4 (full W4A4) body plus the matching gemma4_assistant MTP draft checkpoint in assistant/, so a single hf download gives you everything vllm serve --speculative-config needs.

Measured: Japanese 134 tok/s · English 163 tok/s single-stream on 2× RTX PRO 2000 Blackwell 16 GB (vs 108 baseline) — a QAT-origin, abliterated, Japanese-safe Gemma 4 26B-A4B MoE in 17.6 GB + a 0.84 GB draft, in the practical zone on 2–4 entry-level Blackwell cards.

Lineage: google/gemma-4-26B-A4B-it-qat-q4_0-unquantized (QAT q4_0 → bf16) → huihui-ai abliteration → Lna-Lab NVFP4 W4A4 + loss-less 704→768 MoE pad → this bundle, adding google/gemma-4-26B-A4B-it-qat-q4_0-unquantized-assistant (bf16, unmodified) as the speculative draft.

Why MTP, and why a bundle

Gemma 4's multi-token-prediction is not a head baked into the main checkpoint (Qwen3.6-style). Google ships it as a separate assistant checkpoint that vLLM's --speculative-config loads alongside the target. That means spec-decode normally costs you a second hf download and a path dance. This repo ends that: the assistant lives in assistant/ and you point the config at the local subfolder.

Huihui-gemma-4-26B-A4B-it-qat-abliterated-MTP-NVFP4/
├── model.safetensors          # NVFP4 W4A4 body (17.6 GB)
├── config.json / generation_config.json / processor_config.json
├── tokenizer.json / tokenizer_config.json / chat_template.jinja
├── recipe.yaml                # llm-compressor recipe
└── assistant/                 # gemma4_mtp draft (bf16, 0.84 GB)
    ├── model.safetensors
    ├── config.json            # model_type: gemma4_assistant
    └── tokenizer / chat_template

Quickstart

hf download sakamakismile/Huihui-gemma-4-26B-A4B-it-qat-abliterated-MTP-NVFP4 \
  --local-dir gemma4-26b-mtp
DIR=$(realpath gemma4-26b-mtp)

NCCL_P2P_DISABLE=1 vllm serve "$DIR" \
  --served-model-name gemma4-26b-qat-mtp \
  --tensor-parallel-size 2 \
  --disable-custom-all-reduce \
  --kv-cache-dtype fp8 \
  --max-model-len 16384 \
  --gpu-memory-utilization 0.92 \
  --max-num-batched-tokens 8192 \
  --limit-mm-per-prompt '{"image":0}' \
  --speculative-config "{\"method\":\"gemma4_mtp\",\"model\":\"$DIR/assistant\",\"num_speculative_tokens\":4}"

• The point: model in --speculative-config is the bundled local path — no second download, no HF resolution at serve time.
• TP=2 (2× 16 GB) recommended. Weights ~16.4 GiB don't fit one 16 GB card; TP=4 buys only +6% single-stream on this MoE — spend extra GPUs on a second replica.
• NCCL_P2P_DISABLE=1 + --disable-custom-all-reduce are required on PCIe no-NVLink boxes (TP hangs without them); drop both if you have NVLink/P2P.
• --kv-cache-dtype fp8 doubles KV; with the assistant on board this config still holds maxlen 16384 / KV 23,560 tok at GMU 0.92. Keep CUDA graphs ON (no --enforce-eager).
• vLLM 0.21's quantization-inheritance trap does not fire here: with an explicit draft model path the draft's own config decides (bf16). Only the model:null MTP-from-target path inherits target quantization.

Measured (RTX PRO 2000 Blackwell 16 GB ×2, TP=2, PCIe no-NVLink, fp8 KV, vLLM 0.21.0, 2026-06-12)

Single-stream, T=0 chat completions, ×3 each. Acceptance = accepted/drafted from /metrics.

JA +12–23%, EN +32–51% over baseline. Two lessons paid for in benchmarks:

• EAGLE-3 collapses on Japanese — the English-Magpie-trained draft gets 3–4% JA acceptance and lands below baseline (0.69×). The google MTP assistant holds 35–44% JA acceptance because its 4-layer head re-uses the target's own hidden states instead of imitating its distribution from scratch. If your traffic is non-English, MTP is the only one of the three that pays.
• ngram never pays for itself on free-form chat in either language.

The MoE body already runs 108 tok/s, so MTP's gain here is "fast → faster" (1.2–1.5×). On the dense 31B sibling the same assistant nearly triples throughput — see Huihui-gemma-4-31B-it-qat-abliterated-MTP-NVFP4.

Concurrent (aggregate throughput)

4 / 8 concurrent streams × 256 tok each (T=0, diverse prompts, prefix-cache busted, ×3 averaged). Baseline = same body, no spec-decode (measured with JA prompts; baseline JA≈EN single-stream).

On this MoE, MTP pays up to ~4 concurrent streams; at 8 it is a wash (−2%, run-noise territory). Acceptance stays flat under batch — the shrinking gain isn't the drafter failing, it's the GPUs reaching compute saturation (571 tok/s aggregate), where verifying rejected draft tokens competes with real batch work instead of filling decode bubbles. Worst case is break-even, so keeping MTP resident costs ~nothing at peak load and buys 1.2–1.5× whenever concurrency drops. (The dense 31B sibling, still verification-hungry at 8 streams, keeps +27% JA / +53% EN there.)

The body: QAT × NVFP4 (the finding, in short)

Full W4A4 NVFP4 breaks non-QAT gemma-4 on Japanese long-form — the non-QAT 26B sibling intermittently collapses into repetition loops past ~500 tokens. This QAT-origin body does not: adversarial 1300+-token Japanese essays run to natural EOS with zero loop-detector hits, while English (HumanEval-class) is unaffected. Pattern holds across dense (31B/12B) and MoE: if you want gemma-4 in NVFP4 W4A4, go through a QAT checkpoint — the q4_0-shaped weight distribution is the prior FP4 wants. Full evidence, bake recipe (pure-CPU calibration through the multimodal processor, MoECalibrationModule, pad768 surgery) in the non-MTP card.

Notes

• Abliterated (uncensored). Refusal behavior removed upstream — you are responsible for your deployment. Use responsibly and lawfully.
• NVFP4 is Blackwell-specific; the body will not run on Ampere/Hopper. The bf16 assistant inherits the body's GPU anyway.
• The assistant/ checkpoint is google's, redistributed unmodified under the same Gemma terms; its original model card is included as assistant/README.md.
• Gemma is provided under and subject to the Gemma Terms of Use.

Credits

• Original model & MTP assistant: Google DeepMind (Gemma 4, QAT q4_0)
• QAT-unquantize & abliteration: huihui-ai
• NVFP4 quantization, pad768 surgery, spec-decode measurement & bundle: Lna-Lab · Tooling: llm-compressor / vLLM

Support the Base Model Author (huihui-ai)

If you find the abliterated base useful, please support huihui-ai:

• Ko-fi: https://ko-fi.com/huihuiai
• Bitcoin: bc1qqnkhuchxw0zqjh2ku3lu4hq45hc6gy84uk70ge