	Details
Method	W8A8 FP8 Dynamic
Weights	FP8 (E4M3), per-channel, symmetric
Activations	FP8 (E4M3), dynamic per-token, symmetric
Ignored layers	`lm_head`
Format	`compressed-tensors`
Calibration data	None required (PTQ)

About the Base Model

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled is a reasoning-focused model fine-tuned from Qwen3.5-27B using Chain-of-Thought (CoT) distillation from Claude 4.6 Opus. It uses <think> tags for structured internal reasoning before providing answers.

Key capabilities:

Modular & structured thinking in <think> blocks
Tool calling and coding agent support (Claude Code, OpenCode)
262K token context length
Native developer role support

For full details on the base model, see the original model card.

Usage

With vLLM (recommended)

python

from vllm import LLM, SamplingParams

model = LLM(model="BarraHome/Qwen3.5-27B-Claude-4.6-Opus-FP8-Dynamic")
sampling_params = SamplingParams(max_tokens=2048, temperature=0.6)
output = model.generate(["Hello, tell me about yourself"], sampling_params)
print(output[0].outputs[0].text)

With Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "BarraHome/Qwen3.5-27B-Claude-4.6-Opus-FP8-Dynamic",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "BarraHome/Qwen3.5-27B-Claude-4.6-Opus-FP8-Dynamic",
    trust_remote_code=True,
)

input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))

How It Was Made

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier

model = AutoModelForCausalLM.from_pretrained(
    "Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled",
    dtype="auto",
    low_cpu_mem_usage=True,
    trust_remote_code=True,
)

recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_DYNAMIC",
    ignore=["lm_head"],
)

oneshot(model=model, recipe=recipe, output_dir="Qwen3.5-27B-Claude-4.6-Opus-FP8-Dynamic")

Credits

Base model: Jackrong for the original Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
Quantization: BarraHome using LLM Compressor