license: apache-2.0
base_model:

• SC117/Sage-A1.5-abliterated-MTP-FIT-GGUF
• InternScience/Sage-A1.5
  base_model_relation: quantized
  tags:
• gguf
• strata
• qwen35moe
• moe
• abliterated
• uncensored
• mtp
• vision

Sage-A1.5-abliterated · Strata Q4_K_M GGUF

A Q4_K_M GGUF of SC117's abliterated Sage-A1.5, packed so it loads on the optimized Qwen3.6 engine of
Strata (the q36-r2 branch).
MTP layer and vision encoder included. It is a drop-in replacement for
ykarout/Strata-Sage-A1.5-GGUF, with the abliterated weights.

Safety notice. The source model has had its refusal direction ablated and has no reliable built-in guardrails.
It may comply with harmful, illegal or unsafe requests. Use it only where you provide your own moderation,
access control and legal review.

Built on

This repository only re-packages other people's work:

What was done

SC117's FIT tiers do not load in Strata's Qwen3.6 engine as they are: the engine needs ssm_alpha / ssm_beta in
F32 / BF16 / F16 / Q8_0 (the FIT tiers store them quantized), and it does not take Q3_K routed experts. So the
abliterated BF16 was re-quantized into the same layout as ykarout's file:

llama-quantize --imatrix Sage-A1.5-abliterated-MTP-BF16-imatrix.gguf \
  --tensor-type 'ssm_alpha\.weight=bf16' --tensor-type 'ssm_beta\.weight=bf16' \
  Sage-A1.5-abliterated-MTP-BF16.gguf Sage-A1.5-abliterated-MTP-Q4_K_M-ssmbf16.gguf Q4_K_M 16

• Input BF16 checked against SC117's published sha256 (afbd044f…) before quantizing.
• The tensor types of the result are identical, tensor by tensor, to ykarout's Sage-A1.5-Q4_K_M-ssmbf16.gguf
  (routed experts Q4_K gate / up, Q6_K or Q4_K down; ssm_alpha / ssm_beta BF16; MTP layer kept).
• No other change to the weights or metadata.

A smaller IQ3_S-expert variant (15.75 GiB) was also tried and dropped: it decoded slower on the test machine
(87–94 vs 129–152 tok/s), because the i-quant codebook lookups cost more than the extra cached experts saved.

Files

Running it

Build Strata's q36-r2 branch as in ykarout's card, generate the tokenizer with
python tools/strata_tokenizer.py --gguf <model>.gguf --out <dir>, then use a config like this one
(paths are examples):

{
  "exe": "Strata-sage/build/strata.exe",
  "args": ["--native", "models/sage-abl/Sage-A1.5-abliterated-MTP-Q4_K_M-ssmbf16.gguf",
           "--max-context", "262144", "--prefill", "auto", "--kv", "int8", "--pool-workers", "7",
           "--gqa-attn", "--hybrid-zero-copy", "--expert-prefetch", "2",
           "--expert-profile", "models/sage-abl/expert-profile-agent-v2.prof", "--decode-swaps", "64",
           "--snapshots", "2", "--mtp", "2", "--spec-min-p", "0.5", "--mtp-sampled", "q",
           "--draft-vocab", "16384", "--draft-ctx", "4096", "--vision"],
  "tokenizer": "models/sage-abl/tokenizer",
  "model_name": "sage-a1.5",
  "arch": "qwen35moe",
  "splice_turns": "parsed",
  "vision": {"exe": "engine/strata-vision.exe",
             "mmproj": "models/sage-abl/vision/mmproj-Sage-A1.5-abliterated-bf16.gguf",
             "model": "models/sage-abl/Sage-A1.5-abliterated-MTP-Q4_K_M-ssmbf16.gguf",
             "gpu": false, "max_tokens": 300, "threads": 8},
  "sampling": {"temperature": 1.0, "top_p": 0.95, "top_k": 20, "min_p": 0.0, "presence_penalty": 1.5}
}

--draft-vocab 16384 --draft-ctx 4096 is opt-in: faster MTP on English and code, slower on other scripts per the
q36-r2 notes. Drop --vision and the vision block for text only.

Measured

RTX 5080 16 GB, Ryzen 7 7700 (8 cores), 32 GB RAM, 262K context, int8 KV, MTP 2, --draft-vocab 16384.
Four fixed prompts (three code, one Traditional Chinese), 384 tokens, temperature 0.7, thinking off, warm cache:

90 of 256 experts per layer in VRAM, MTP draft acceptance ~70–78%, the same as the original file.
Vision (CPU encoder): a 640×400 test image with two shapes and the text "STRATA 7419" was described correctly,
about 1.7 s for 290 image tokens.

Not measured

• KL / perplexity of this quantization against the BF16 (SC117's numbers are for their own tiers, not this file).
• Refusal rate after quantization.
• Long-context speed (above ~17K tokens) and the vision encoder's effect on decode speed.

License

Apache-2.0, inherited from Sage-A1.5. All credit for the model to InternScience / Shanghai AI Laboratory, for the
abliteration to SC117, and for the Strata packing and engine work to ykarout / xyehya and the Strata authors.