Sage-A1.5 (Q4_K_M, uncensored)

A locally-runnable, refusal-suppressed build of the Qwen3.5-MoE family (256 routed experts, 35B-A3B class), packaged as a single Q4_K_M GGUF plus an F16 vision projector.

• This repo (Hugging Face): single-file GGUF, easiest download
• GitHub mirror (split parts + scripts): https://github.com/zimuge-doudou/sage-a1.5

File	Size	What
sage-a1.5-q4km-abl4.0.gguf	27.56 GB	Abliteration already merged into the weights (LoRA scale 4.0 baked in), 753 tensors
sage-a1.5-mmproj.gguf	899 MB	F16 vision projector

⚠️ The plain Q4_K_M base (sage-a1.5-q4km.gguf) is not shipped: that file is the stock quantization and refuses like the original model. Download the abl4.0 file above.

• Base architecture: qwen35moe (Qwen3.5-MoE) — Apache-2.0
• Context: trained/exported for long context; we run 73,728 tokens with q8_0 KV cache
• Serving: llama.cpp llama-server (OpenAI-compatible /v1)
• No adapter needed at serve time: -m sage-a1.5-q4km-abl4.0.gguf is enough.

What we changed

1. Refusal-direction abliteration on the post-training chat model (no fine-tuning on harmful data, no reward model): the refusal direction is projected out of the residual stream.
2. The edit is fused into the shipped weights. We first produced a rank-1 abliteration adapter, found scale 4.0 to be the sweet spot (higher starts to degrade reasoning), then merged it into the GGUF with llama-export-lora --lora-scaled …:4.0. Loading the file needs no --lora flag.
3. No safety layer, no system-prompt filtering is added by us.

Verification we ran on the merged file itself (not on the adapter path): 5 policy-adjacent prompts → 0 refusals, identical to the un-merged serving setup, and the coding bench still passes 3/3.

The result keeps the base model's coding/reasoning ability while dropping most stock refusals. It is not a new pretrain and it is not aligned — see Limitations.

Benchmarks (our own smoke tests, not a leaderboard submission)

Measured 2026-10-23 on a MacBook Pro M4 Pro / 64 GB, llama.cpp, temp=0.4, enable_thinking=false, same prompts for every model we compared.

Test	Result
Coding smoke bench (3 tasks: positional arithmetic, branchless max, small Python driver)	3 / 3
Refusal probe (5 policy-adjacent prompts)	0 / 5 refusals
Decode speed (Q4_K_M, M4 Pro, single stream)	~50 tok/s
OpenAI tools / function calling	supported (finish_reason=tool_calls)
Same bench on Qwen3.8-27B abliterated (Q4)	3 / 3, ~11 tok/s
Same bench on Qwen3.5-27B abliterated (Q4)	2 / 3, ~56 tok/s

Raw per-task notes: https://github.com/zimuge-doudou/sage-a1.5/blob/main/docs/eval.md

Download

# whole file, resumable
hf download amy-sage/sage-a1.5 sage-a1.5-q4km-abl4.0.gguf --local-dir .
hf download amy-sage/sage-a1.5 sage-a1.5-mmproj.gguf --local-dir .

GitHub (split into 15 parts, 14 × 1.8 GiB + one 1.06 GiB, with SHA256SUMS):

gh release download v1.0 -R zimuge-doudou/sage-a1.5 -D ./parts
cd ./parts && cat sage-a1.5-q4km-abl4.0.gguf.part-{a..o} > sage-a1.5-q4km-abl4.0.gguf
sha256sum -c ./parts/SHA256SUMS      # macOS: shasum -a 256 -c

Run

llama-server \
  -m sage-a1.5-q4km-abl4.0.gguf \
  --mmproj sage-a1.5-mmproj.gguf \
  -c 73728 -ctk q8_0 -ctv q8_0 -ngl 999 -fa on \
  --host 127.0.0.1 --port 8090

OpenAI-compatible client:

from openai import OpenAI
c = OpenAI(base_url="http://127.0.0.1:8090/v1", api_key="local")
print(c.chat.completions.create(
    model="sage-a1.5",
    messages=[{"role": "user", "content": "Explain MoE routing in two sentences."}],
    max_tokens=512,
).choices[0].message.content)

Hardware notes

• 25.7 GiB of weights: needs ≥ 32 GB unified memory to load comfortably at 8K context; 64K context with q8_0 KV needs roughly 40 GB.
• On Apple silicon, -ngl 999 --flash-attn on puts everything on the GPU.
• On a 24 GB machine, use -c 8192 and expect slower prompt processing.

Limitations

• Uncensored by construction. It will answer things the base model refuses. That is the point, but it also means it will happily produce wrong, offensive, or unsafe text — including confident hallucinations. It has no truthfulness guarantee.
• Numbers are the weak spot. We measured it inventing disk-usage figures when asked about the host machine; feed it real values (or a tool result) instead of trusting recall.
• Self-reports are unreliable — treat any "I did X" from this model as a claim to verify.
• Not evaluated on safety benchmarks; not for user-facing deployment without your own guardrails.

License

Base model and this derivative: Apache-2.0 (see LICENSE). You must keep the Qwen attribution when redistributing.

Citation

Sage-A1.5 (2026). Abliterated Q4_K_M build of Qwen3.5-MoE, Apache-2.0.
https://huggingface.co/amy-sage/sage-a1.5
https://github.com/zimuge-doudou/sage-a1.5