library_name: transformers
license: apache-2.0
base_model: sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored
base_model_relation: quantized
pipeline_tag: image-text-to-text
tags:
- heretic
- abliterated
- uncensored
- awq
- w4a16
- compressed-tensors
- moe
- vision
- agentic
- tool-use
Sage-A1.5 Heretic Abliterated Uncensored, W4A16
W4A16 (int4 weights, group size 32) quantization of sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored, which is a Heretic abliteration of InternScience/Sage-A1.5. No extra fine-tuning. Architecture, tokenizer, chat template, context length and vision tower are unchanged.
Quantization
- Format: compressed-tensors,
pack-quantized, int4, asymmetric, group size 32. Data-free (no calibration),memoryless-mseobserver, via llm-compressormodel_free_ptq. - Quantized: routed experts in layers 1-38 (29,184 matrices).
- Kept in BF16: attention, linear attention, shared expert, router, layers 0 and 39, embeddings,
lm_head, vision tower. - Size: 26.7 GB, 15 shards. No MTP weights (the source has none).
Usage
Tested on vLLM 0.32.0, 2x RTX 3090, tensor parallel 2, 262,144 context.
vllm serve <this-repo> \
--tensor-parallel-size 2 \
--max-model-len 262144 \
--kv-cache-dtype fp8_e5m2 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder
Suggested sampling (from Sage-A1.5): temperature 1.0, top_p 0.95, top_k 20, min_p 0.0, presence_penalty 1.5.
Notes
Measured on the author's rig, small samples:
- Refusals: 0 of 12 on a keyword-screened probe set.
- Distance to the BF16 source on 48 benign prompts: mean KL 0.0072, top-1 agreement 97.8% (top-20 approximation).
- Abliteration costs instruction-hierarchy robustness. On PromptAuthority-15 this model scores 38.0 against 66.7 for the censored Sage-A1.5 AWQ-INT4, and it can follow instructions injected into fetched or tool content. Other packs: ToolCall-15 90.3, InstructFollow-15 97.0, ReasonMath-15 76.7, HermesAgent-20 73.7 (censored AWQ: 96.7 / 94.3 / 72.7 / 78.0).
- Do not give it destructive or outbound tools on untrusted content without confirmation steps.
This is an uncensored model. You are responsible for how you use it.
License
Apache-2.0, inherited from Sage-A1.5. The Heretic tool is AGPL-3.0.