Parameter	Value
Method	FP8 E4M3 block-wise
Block size	128x128
Scale dtype	BF16
Scale naming	`weight_scale_inv`
Activation scheme	Dynamic
Original size	~66 GB (BF16)
Quantized size	~35 GB

What's quantized

All linear projections in attention (q/k/v/o_proj), GatedDeltaNet (in_proj_qkv, in_proj_z, out_proj), and MLP experts (gate/up/down_proj for all 256 experts + shared expert).

What's kept in BF16

Embeddings, lm_head, norms, MoE router gates, GDN precision-sensitive params (conv1d, in_proj_a, in_proj_b), vision tower.

Usage with SGLang

bash

CUDA_DEVICE_ORDER=PCI_BUS_ID \
CUDA_VISIBLE_DEVICES=0 \
SGLANG_ENABLE_JIT_DEEPGEMM=0 \
SGLANG_ENABLE_SPEC_V2=1 \
  python -m sglang.launch_server \
    --model-path nivvis/Qwen3.5-35B-A3B-heretic-v2-FP8 \
    --tool-call-parser qwen3_coder \
    --port 30000 --host 0.0.0.0 \
    --mem-fraction-static 0.85 \
    --context-length 32768 \
    --attention-backend triton \
    --reasoning-parser qwen3 \
    --mamba-scheduler-strategy extra_buffer \
    --trust-remote-code

Usage with vLLM

python

from vllm import LLM
model = LLM("nivvis/Qwen3.5-35B-A3B-heretic-v2-FP8", trust_remote_code=True)

Performance

Tested on NVIDIA RTX PRO 6000 Blackwell (98 GB):

Decode throughput: ~134 tokens/s (single request)
VRAM usage: ~34 GB model + KV cache
Default FP8 kernel configs (no device-specific tuning)

Notes

No MTP: Heretic abliteration strips MTP weights. Do not use --speculative-algo NEXTN (accept rate ~0.25). Restoring MTP by copying the head from the original Qwen model and fine-tuning against heretic hidden states is a potential future improvement.
Thinking model: Supports <think> reasoning. Use "chat_template_kwargs": {"enable_thinking": false} to disable.
Validated against official Qwen/Qwen3.5-35B-A3B-FP8: 0 missing scale tensors, 0 unexpected tensors, matching quantization config.

Credits

Base model: Qwen/Qwen3.5-35B-A3B by Qwen Team
Abliteration: llmfan46/Qwen3.5-35B-A3B-heretic-v2 using Heretic
FP8 quantization: nivvis