Ektome-Qwen3.8-27B-PristinelyUncensored

GGUF status — 2026-08-15

Q4_K_M has been rebuilt and re-uploaded. It is correct. If you downloaded any GGUF from this repo before 2026-08-15 01:00 UTC, delete it and pull again.

The earlier files failed with error loading model: check_tensor in llama.cpp / LM Studio, with MTP on or off. The header declared block_count=65 / nextn_predict_layers=1 while the file shipped only blk.0..63 and no MTP tensors — the header disagreed with its own contents, so the block-count check failed before MTP was ever consulted. Root cause was upstream: the multi-token-prediction head was lost when the abliterated model was saved, so metadata (read from config.json) and tensors disagreed.

The rebuilt Q4_K_M is verified: header block_count=65 matches 65 actual blocks, blk.64 (the MTP block) is present, and it loads and generates.

The bf16 safetensors were never affected — 1199 tensors, exact parity with Qwen/Qwen3.8-27B.

Reported and correctly diagnosed by MrGuzzy and Scottbyrd. Thank you both.

Quantized siblings

Which one do you want?

• Long context, and you'll install a fork → HOMEUSER-16-24. It is the only build that quantizes embed_tokens and lm_head — two 248320 x 5120 tensors every other release leaves in bf16 — which frees ~6 GB for KV cache and is why 262,144 tokens fit on a 24 GB card. Measured on eight GPUs (5090 / 4090 / 3090 Ti / 3090 / A4000 / 5060 Ti / 5070 Ti / 4080 Super); MMLU 0.7995, HumanEval 0.8902, MMStar 0.6633. Stock vLLM cannot bind those two tensors, hence the fork.
• You want unmodified vLLM or SGLang → GPTQ-MTP. 72.6 tok/s at batch 1 / 512.1 at batch 8 on an RTX PRO 6000 (sm_120), vLLM 0.27.1. It cannot be loaded by transformers — the [48, 5120] DeltaNet projections are excluded from Marlin via GPTQModel's dynamic config, which only vLLM and SGLang understand. See that repo's card for the two load requirements.
• llama.cpp / LM Studio / Ollama → the GGUF.

MTP (multi-token prediction) — present, and off unless you enable it

These weights include the complete 15-tensor MTP head, byte-identical to Qwen/Qwen3.8-27B (Ektomē edits the language decoder only). Speculative decoding is not automatic — vLLM ignores the head unless you ask for it:

from vllm import LLM
llm = LLM(
    model="Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored",
    speculative_config={"method": "qwen3_5_mtp", "num_speculative_tokens": 1},
)

vllm serve Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored   --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":1}'

Measured on the 4-bit GPTQ sibling (RTX PRO 6000, vLLM 0.27.1, greedy): +35% at batch 1, +30% at batch 8, and +63% on long-form code generation — the gain is largest on structured output, because the draft head is accepted more often there.

⚠️ Those figures are from the 4-bit build, not from these bf16 weights. bf16 inference is more compute-bound, so the speedup here may differ — we have not measured it. The mechanism and the invocation are identical; only the magnitude is unverified for bf16.

One caveat that applies to both: at greedy decoding, speculative and non-speculative output are each individually deterministic but differ from one another. Pick one configuration and stay on it if you cache completions or run evaluation harnesses.

What this bf16 contains

• 1199 tensors — exact parity with Qwen/Qwen3.8-27B, including the 15-tensor MTP (multi-token-prediction) head and the full vision tower. Verified tensor-by-tensor against base.
• 128 residual-write matrices excised (o_proj / out_proj / down_proj), language decoder only. The MTP head, the vision tower, and the 2 MTP-block matrices that pass a naive dim check were all left untouched.
• Ektomē receipts: refusal compliance 0.000 → 1.000 (n=100), MMLU-val 0.812 → 0.818 (Δ +0.005 against SE 0.020 — capability held, not improved), generative gate clean (foreign 0.000, degeneration 0.000, instruction-following 0.400 → 0.400). Winner A:frac=0.6.
• This is the parent artifact: every quantized sibling is derived from these weights, so any fine-tune, re-quantization or re-abliteration should start here rather than from a GGUF.

The receipt — pristine vs Ektomē (before / after)

Fully uncensored, capability +0.005 vs pristine. Capability = MMLU-validation loglikelihood accuracy; refusal on harmful = AdvBench, judge-free keyword classifier, thinking disabled so chain-of-thought can't mask the answer.

How we know it's real — the catcher's verdict

Every claim here is gated by a surrogate-null catcher, not asserted. The excision is applied to pristine and measured; a capability change counts only if it beats the MMLU sampling floor.

$ ektome_abliterate  Qwen/Qwen3.8-27B
  PRISTINE (before): compliance=0.000   mmlu-val=0.812
  EKTOME   (after):  compliance=1.000   mmlu-val=0.818   (dcap=+0.005)
  -> uncensored AND capability preserved, zero training.

Why it's different

• Uncensored — complies on harmful prompts the base refuses.
• Capability-preserved — capability stays close to the pristine base, not the capability-lower crude-abliterated models most tools ship.
• Zero training — a single surgical weight edit on pristine, seconds of compute.
• Honest — every claim is gated by a surrogate-null test, not marketing.

Run it locally

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored")
model = AutoModelForCausalLM.from_pretrained("Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored", torch_dtype="bfloat16", device_map="auto")

GGUF (any gaming PC): Q4_K_M (16.8 GB) — verified to load in llama.cpp / LM Studio / Ollama. 4-bit GPTQ/Marlin (17.7 GB, vLLM or SGLang only) is available as a separate repo and is the fastest option — see its card for measured throughput. LM Studio / Ollama.

The Ektomē line & feedback

If this is useful, a ❤️ like helps others find it. It's part of a line — the same catcher-gated surgery across frontier small models (Qwen, Mistral, Llama, Nemotron) and larger bases. Want a specific base next or have head-to-head numbers? Open a Discussion — the line is steered by what people actually run.

Honest notes

• Uncensored: this model follows instructions the base refuses. Use it lawfully and responsibly; you are accountable for what you generate.
• Compliance uses a judge-free keyword classifier (thinking disabled); the capability delta is gated against a surrogate-null noise floor.
• Method: produced by Ektomē, a capability-preserving surgical-abliteration method. Results are reported; the full recipe is not (yet).

License

Inherits the base model's license (Qwen3.8-27B → apache-2.0). No warranty; your use, your responsibility.