license: other
license_name: polyform-small-business-1.0.0
license_link: LICENSE
base_model: IstroSec/ThinkingCap-Qwen3.8-27B-abliterated
base_model_relation: quantized
library_name: transformers
pipeline_tag: image-text-to-text
language:

• en
• multilingual
  tags:
• qwen3_5
• qwen3_8
• thinkingcap
• abliterated
• uncensored
• fp8
• fp8_dynamic
• compressed-tensors
• llmcompressor
• vllm
• mtp
• vision

ThinkingCap-Qwen3.8-27B-abliterated-FP8-DYNAMIC

FP8 (W8A8, dynamic per-token activations) quantization of ThinkingCap-Qwen3.8-27B-abliterated, the uncensored variant of bottlecapai/ThinkingCap-Qwen3.8-27B.

36.8 GB (from 55.6 GB bf16). That's ~6 GB more than an FP8 build that also quantizes the DeltaNet block; see below for why it doesn't. Near-lossless. Serves on FP8-capable GPUs (Hopper, Ada, Blackwell, DGX Spark) natively, and on Ampere through vLLM's Marlin weight-only FP8 fallback — no custom kernels either way.

What is quantized

Produced with llm-compressor, scheme FP8_DYNAMIC: per-channel FP8 (E4M3) weights, per-token dynamic FP8 activations. Data-free — no calibration set.

Quantized: all Linear modules in the 64 decoder layers' MLPs and the 16 full-attention layers' q/k/v/o_proj.

Kept in bf16 on purpose:

The MTP Linears are listed in quantization_config.ignore so vLLM loads them as bf16 rather than looking for scales that don't exist.

Evaluation

Method: Heretic --evaluate-model against the bf16 abliterated checkpoint, non-thinking mode. KL under 0.01 is imperceptible; under 0.05 is good.

Chain of provenance for the full picture: original ThinkingCap refuses 97/100; the bf16 abliteration brings that to 6/100 at KL 0.065 vs. original; this quantization adds the KL above on top of that.

Usage

vLLM (recent release; cu130 wheels on DGX Spark)

vllm serve IstroSec/ThinkingCap-Qwen3.8-27B-abliterated-FP8-DYNAMIC \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
  --max-model-len 65536 --gpu-memory-utilization 0.85

KV cache: only the 16 full-attention layers have one, so the FP8 KV cache (--kv-cache-dtype fp8) buys less than on a dense model. Leave it at auto unless you're memory-bound; if you do enable it, fp8_e5m2 needs no scales.

Sampling (Qwen3.8 recommendations, which ThinkingCap uses unchanged): thinking mode temperature 1.0, top_p 0.95, top_k 20, min_p 0; non-thinking mode temperature 0.7, top_p 0.8, top_k 20, presence_penalty 1.5. Thinking budget via chat_template_kwargs: {"reasoning_effort": "xhigh"} — xhigh (default, recommended), medium, or low.

Transformers loads it too (dequantized to bf16 on load, so it needs the full ~56 GB):

from transformers import AutoModelForImageTextToText
m = AutoModelForImageTextToText.from_pretrained("IstroSec/ThinkingCap-Qwen3.8-27B-abliterated-FP8-DYNAMIC", device_map="cuda")

Not for llama.cpp — this is compressed-tensors format. GGUF builds are made separately from the bf16 source.

Reproduce

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
recipe = QuantizationModifier(
    targets="Linear", scheme="FP8_DYNAMIC",
    ignore=["lm_head", "re:.*visual.*", "re:.*linear_attn.*"],
)
oneshot(model=model, recipe=recipe)
# then copy mtp.* from the bf16 checkpoint and add the MTP Linears to quantization_config.ignore

Limitations

Everything from the bf16 card applies: no safety filter, 6/100 residual refusals, thinking mode not separately evaluated. You are the safety layer.

License

PolyForm Small Business License 1.0.0 + BottleCap personal-use grant, inherited from ThinkingCap (see LICENSE). Upstream Qwen materials and the abliteration adapter are Apache-2.0 (see NOTICE). Commercial use beyond the PolyForm terms: contact BottleCap AI.

Credits

bottlecapai (ThinkingCap) · MuXodious (abliteration adapter) · p-e-w/heretic · vllm-project/llm-compressor · Qwen team