license: mit
base_model:

• huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated
  language:
• en
• ja
  library_name: gguf
  pipeline_tag: text-generation
  tags:
• gguf
• llama.cpp
• quantized
• iq3_m
• gsq-rco
• abliterated
• uncensored
• qwen3_5_moe
• moe
• mtp

Ornith-1.5-35B-A3B-abliterated-GSQ-RCO-IQ3_M

A GSQ-RCO-style IQ3_M GGUF build of the abliterated Ornith-1.5-35B-A3B
(huihui-ai), combining locally converted weights with the per-tensor
quantization allocation of the community GSQ-RCO build:

• Weights: huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated, converted to GGUF (Q8_0) locally with llama.cpp and used as the quantization input
• Allocation donor (body, blocks 0–39): pfeifferj/Ornith-1.5-35B-A3B-GSQ-RCO-GGUF (its 3-bit build)
• Allocation donor (MTP block 40): the APEX GSQ-RCO IQ3_XXS build by lavendercomputing (the pfeifferj build excludes the MTP block; its per-tensor types for block 40 were measured from the APEX file's GGUF header)
• Base model: ornith-ai/Ornith-1.5-35B-A3B by DeepReinforce AI (Ornith AI), MIT license

⚠️ This model is abliterated. Its safety alignment was deliberately
removed upstream. It has no guardrails. You are responsible for how you
use it.

Files

Unlike the 9B heretic GGUF conversions, this file retains the MTP / nextn
block (block 40, 20 tensors, qwen35moe.nextn_predict_layers = 1), so
llama.cpp speculative decoding via --spec-type draft-mtp can be used.

For image input you additionally need the vision projector (mmproj),
which is not included here.

How it was made

This is an allocation-transfer build, in the same style as most
community "GSQ-RCO" GGUFs — it is not a from-scratch GSQ/RCO
recomputation:

1. The per-tensor quantization types were extracted from the donors' GGUF
   headers (733 body tensors from pfeifferj 3-bit + 20 MTP tensors from the
   APEX build = 753).
2. An importance matrix was generated locally from the abliterated weights
   with llama-imatrix, using a mixed calibration set: English wikitext-2
   • Japanese Wikipedia (20231101), 200 chunks × 512 context.
3. llama-quantize --allow-requantize --imatrix … --tensor-type-file …
   was run with the Q8_0 conversion of the huihui weights as input.

Toolchain: a prebuilt llama.cpp set (0.4.1-dev, build 10976) used for all
steps (imatrix, quantize, verification) on a desktop CPU (Ryzen 5 5500).
The desktop has no C++ toolchain, so the pinned b11937 source tree could
not be built there; correctness of the allocation was instead verified
exhaustively after quantization (see below).

Allocation fidelity — read this before comparing with the donors

Measured after the build with a full header diff against the donor tables:
of the 554 controlled (≥2D) tensors, 554/554 carry exactly the donor's
quantization type (100.00%), with 0 missing tensors and 0 downgrades
(no tensor was silently re-typed by quantizer built-in overrides). The
remaining 199 tensors are 1-D (norms, biases, SSM parameters) and follow
llama.cpp defaults, as in the donor builds.

Verification

• Header: 753 tensors, qwen35moe.block_count = 41,
  qwen35moe.nextn_predict_layers = 1 (MTP block retained).
• Smoke tests (llama-completion, CPU, greedy): EN prompt "The capital of
  France is" → "Paris." with a correct continuation; JA prompt
  「日本の首都は」 → 「東京です。」 with a natural continuation.
• Held-out perplexity (same file and settings for both: 16 chunks × 512
  context): this build 12.3454 ± 0.5644 vs the pfeifferj 3-bit donor
  12.7927 ± 0.5911. Note the bases differ (abliterated weights here,
  original weights in the donor), so read this as a product-level
  comparison, not a quantizer-only one.

Usage (llama.cpp)

llama-cli -m Ornith-1.5-35B-A3B-abliterated-GSQ-RCO-IQ3_M.gguf \
  --temp 0.6 --top-p 0.95 --top-k 20 --repeat-penalty 1.05 \
  -p "Hello"

llama-server -m Ornith-1.5-35B-A3B-abliterated-GSQ-RCO-IQ3_M.gguf -c 8192

Sampling guidance inherited from the upstream cards for this model family:
repeat-penalty 1.05 is the sweet spot (1.0 can loop, 1.1 can truncate);
greedy decoding is not recommended. The model reasons inside <think>
blocks by default; use your runtime's reasoning controls if you want that
off.

Credits

• DeepReinforce AI / Ornith AI — the base model, Ornith-1.5-35B-A3B (MIT).
• huihui-ai — the abliterated weights.
• pfeifferj — the GSQ-RCO donor build whose body allocation this file reuses.
• lavendercomputing — the APEX GSQ-RCO build whose MTP-block allocation this file reuses.
• ISTA-DASLab — the GSQ / RCO quantization research this style of build is based on.
• ggml-org / llama.cpp — conversion, imatrix and quantization tooling.

License

MIT. License metadata verified on Hugging Face on 2026-11-04 for the
whole weight lineage — the base model (ornith-ai), the abliterated weights
(huihui-ai), and the body allocation donor (pfeifferj) are all MIT — and
the full MIT text is included in this repository as LICENSE. One
transparency note: the MTP-block (block 40) allocation was measured from
lavendercomputing's APEX build, whose repository displays Apache-2.0,
while the vendor lineage it derives from is MIT; the block-40 types follow
the quantization rule lavendercomputing publishes in their own README
(2-D tensors in Q8_0, norms in F32).