Model Architecture: Gemma 4
- Input: Text / Image
- Output: Text
Model Optimizations:
- Weight quantization: FP4
- Activation quantization: FP4
Release Date: 2026-07-20
Version: 2.0
Quantized by: xdavxd
Base Model: coder3101/gemma-4-26B-A4B-it-heretic
Original Model: google/gemma-4-26B-A4B-it

This model is a quantized version of coder3101/gemma-4-26B-A4B-it-heretic. It was evaluated on several tasks to assess its quality in comparison to the original model.

Model Optimizations

This model was obtained by quantizing the weights and activations of coder3101/gemma-4-26B-A4B-it-heretic to FP4 data type using the NVFP4 format, ready for inference with vLLM. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 64%.

Only the MoE expert MLP weights and activations are quantized to FP4 using NVIDIA Model Optimizer with the nvfp4_experts_only-kv_fp8_cast recipe. Self-attention projections, shared MLP layers, vision tower, embedding, output head, and MoE router layers are kept in their original BF16 precision. KV cache uses FP8 with constant amax at inference time.

v2 uses NVIDIA's experts-only quantization strategy, which quantizes only the MoE expert layers (~90% of parameters) while preserving the quality-sensitive attention and shared MLP layers at full precision. This follows established MoE quantization best practices showing expert layers are significantly more tolerant of low-bit quantization than shared pathway layers.

Deployment

Use with vLLM

This model can be deployed using vLLM. For detailed instructions including multi-GPU deployment, multimodal inference, thinking mode, function calling, and benchmarking, see the Gemma 4 vLLM usage guide.

Start the vLLM server:

vllm serve xdavxd/gemma-4-26B-A4B-it-heretic-v2-NVFP4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90

To enable thinking/reasoning and tool calling:

vllm serve xdavxd/gemma-4-26B-A4B-it-heretic-v2-NVFP4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90 \
  --enable-auto-tool-choice \
  --reasoning-parser gemma4 \
  --tool-call-parser gemma4 \
  --chat-template examples/tool_chat_template_gemma4.jinja \
  --default-chat-template-kwargs '{"enable_thinking": true}' \
  --limit-mm-per-prompt '{"image": 4, "audio": 1}' \
  --async-scheduling

Tip: For text-only workloads, pass --limit-mm-per-prompt '{"image": 0, "audio": 0}' to skip vision encoder memory allocation and free up GPU memory for a longer context window.

Send requests to the server:

python

from openai import OpenAI

openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8001/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

model = "xdavxd/gemma-4-26B-A4B-it-heretic-NVFP4"

messages = [
    {"role": "user", "content": "Explain quantum mechanics clearly and concisely."},
]

outputs = client.chat.completions.create(
    model=model,
    messages=messages,
)

generated_text = outputs.choices[0].message.content
print(generated_text)

Creation

This model was created by applying NVFP4 quantization with NVIDIA Model Optimizer, as presented in the code snippet below.

shell

python3 examples/llm_ptq/hf_ptq.py \
  --model ~/models/gemma-4-26b-a4b-heretic \
  --recipe general/ptq/nvfp4_experts_only-kv_fp8_cast \
  --export_path ~/models/gemma-4-26b-a4b-heretic-nvfp4 \
  --calib_size 1024 \
  --export_fmt hf

Evaluation

This model was evaluated on GSM8K Platinum, MMLU-Pro, IFEval, MATH-500, AIME 2025, GPQA Diamond, and LiveCodeBench v6 using lm-evaluation-harness and lighteval, served with vLLM (OpenAI-compatible API). All evaluations were performed with thinking disabled (chat template default).

Performance

(as reported by coder3101)

Metric	This model	Original model (google/gemma-4-26B-A4B-it)
KL divergence	0.0499	0 (by definition)
Refusals	11/100	100/100

Accuracy

Original and RedHatAI benchmarks taken from RedHatAI, comparison to see how the heretic model converted to NVFP4 stacks up against the non-heretic versions. Recovery vs. Original to see how the combination of heretic ARA (Arbitrary-Rank Ablation) and then quantization affects accuracy.

Category	Benchmark	google/gemma-4-26B-A4B-it	RedHatAI/gemma-4-26B-A4B-it-NVFP4	xdavxd/gemma-4-26B-A4B-it-heretic-NVFP4	xdavxd/gemma-4-26B-A4B-it-heretic-v2-NVFP4	Recovery
Instruction Following	IFEval (0-shot, prompt-level strict)	89.96	87.24	87.19	89.16	99.1%
IFEval (0-shot, inst-level strict)	93.21	91.29	91.05	92.45	99.2%
Reasoning	GSM8K Platinum (0-shot, strict-match)	95.43	94.43	94.49	95.53	100.0%
MMLU-Pro (0-shot, custom-extract)	83.47	81.65	-	81.81*	98.0%
MATH-500 (0-shot, pass@1)	84.80	87.60	85.13	85.93	101.3%
AIME 2025 (0-shot, pass@1)	80.00	66.25	58.75	77.09	96.36%
GPQA Diamond (0-shot, pass@1)	73.20	72.39	66.67	73.57	100.5%
Coding	LiveCodeBench v6 (0-shot, pass@1)	74.48	68.19	-	-	-.-%

^{* MMLU Pro results are 1 run}

Reproduction

The results were obtained using the following commands:

Each benchmark was run 3 times with different judgmental seeds (1234, 2345, 3456) and the scores were averaged; AIME 2025 used 8 seeds.

Ran using a docker image, you will need to tweak these for your own environment. Kept parameters similar to the RedHatAI runs.

vLLM server (all benchmarks):

sudo docker run --rm -it \
  --gpus all \
  --ipc=host \
  --network host \
  -v ~/models/gemma-4-26b-a4b-it-heretic-v2-nvfp4:/models/heretic2:ro \
  -e VLLM_USE_V2_MODEL_RUNNER=1 \
  ghcr.io/timothystewart6/vllm-gb10:latest \
  vllm serve /models/heretic2 \
  --host 0.0.0.0 --port 8001 \
  --served-model-name heretic-v2 \
  --max-model-len 69632 \
  --gpu-memory-utilization 0.50 \
  --enable-auto-tool-choice \
  --reasoning-parser gemma4 \
  --tool-call-parser gemma4 \
  --limit-mm-per-prompt '{"image":0,"audio":0}' \
  --kv-cache-dtype fp8 \
  --max-num-seqs 32 \
  --async-scheduling

GSM8K Platinum (lm-eval, 0-shot, 3 repetitions)

for SEED in 1234 2345 3456; do
  lm_eval --model local-chat-completions \
    --tasks gsm8k_platinum_cot_llama \
    --model_args "model=heretic-v2,max_length=36096,base_url=http://0.0.0.0:8001/v1/chat/completions,num_concurrent=32,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
    --num_fewshot 0 \
    --apply_chat_template \
    --output_path "results_gsm8k_seed${SEED}.json" \
    --seed $SEED \
    --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=64,max_gen_toks=32000,seed=${SEED}"
  echo "Seed $SEED complete"
done

MMLU-Pro (lm-eval, 0-shot, 3 repetitions)

for SEED in 1234 2345 3456; do
  lm-eval run --model local-chat-completions \
    --tasks mmlu_pro \
    --model_args "model=heretic-v2,max_length=36096,base_url=http://0.0.0.0:8001/v1/chat/completions,num_concurrent=32,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
    --num_fewshot 0 \
    --apply_chat_template \
    --output_path "results_mmlu_pro_seed${SEED}.json" \
    --seed $SEED \
    --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=64,max_gen_toks=32000,seed=${SEED}"
  echo "MMLU-Pro seed $SEED complete"
done

IFEval (lm-eval, 0-shot, 3 repetitions)

for SEED in 1234 2345 3456; do
  lm_eval --model local-chat-completions \
    --tasks ifeval \
    --model_args "model=heretic-v2,max_length=36096,base_url=http://0.0.0.0:8001/v1/chat/completions,num_concurrent=32,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
    --num_fewshot 0 \
    --apply_chat_template \
    --output_path "results_ifeval_seed${SEED}.json" \
    --seed $SEED \
    --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=64,max_gen_toks=32000,seed=${SEED}"
  echo "IFEval seed $SEED complete"
done

MATH-500, AIME 2025, GPQA Diamond (lighteval, 3 repetitions; 8 for AIME 2025)

litellm_config.yaml:

yaml

model_parameters:
  provider: hosted_vllm
  model_name: hosted_vllm/heretic-v2
  base_url: http://0.0.0.0:8001/v1
  api_key: ''
  timeout: 7200
  concurrent_requests: 32
  generation_parameters:
    temperature: 1.0
    max_new_tokens: 65536
    top_p: 0.95
    top_k: 64
    seed: 1234

# GPQA Diamond (3 seeds)
for SEED in 1234 2345 3456; do
  sed -i "s/seed: .*/seed: $SEED/" litellm_config.yaml
  lighteval endpoint litellm litellm_config.yaml 'gpqa:diamond|0' \
    --output-dir "results_gpqa_seed${SEED}/" --save-details
  echo "GPQA Diamond seed $SEED complete"
done

# MATH-500 (3 seeds)
for SEED in 1234 2345 3456; do
  sed -i "s/seed: .*/seed: $SEED/" litellm_config.yaml
  lighteval endpoint litellm litellm_config.yaml 'math_500|0' \
    --output-dir "results_math500_seed${SEED}/" --save-details
  echo "MATH-500 seed $SEED complete"
done

# AIME 2025 (8 seeds)
for SEED in 1234 2345 3456 4567 5678 6789 7890 8901; do
  sed -i "s/seed: .*/seed: $SEED/" litellm_config.yaml
  lighteval endpoint litellm litellm_config.yaml 'aime25|0' \
    --output-dir "results_aime_seed${SEED}/" --save-details
  echo "AIME 2025 seed $SEED complete"
done

LiveCodeBench v6 (lighteval, 3 repetitions)

litellm_config.yaml:

yaml

model_parameters:
  provider: hosted_vllm
  model_name: hosted_vllm/heretic
  base_url: http://0.0.0.0:8001/v1
  api_key: ''
  timeout: 7200
  concurrent_requests: 32
  generation_parameters:
    temperature: 1.0
    max_new_tokens: 32768
    top_p: 0.95
    top_k: 64
    seed: 1234

for SEED in 1234 2345 3456; do
  sed -i "s/seed: .*/seed: $SEED/" litellm_config.yaml
  lighteval endpoint litellm litellm_config.yaml 'lcb:codegeneration_v6|0' \
    --output-dir "results_lcb_seed${SEED}/" --save-details
  echo "LiveCodeBench v6 seed $SEED complete"
done