Qwen3.8-Flash-Next-GSQ-RCO-Coder-Uncen · IQ1_M · 256 experts/layer (GGUF)

The Coder, uncensored. This release takes ISTA-DASLab's pruned GSQ-RCO Coder (IQ1_M tier, 256/512 experts per layer) and swaps its refusal behavior for the abliterated (uncensored) core of SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF — as a weights-only transplant. No runtime control vector is needed: the abliteration is in the tensors. ("Uncen" here means exactly that abliterated core; the original build was prepared under the name Coder-Abliterated, which MANIFEST.json keeps.)

• Base weights: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF @ 5348543e (IQ1_M tier — the Coder's own 256-experts/layer pruning and expert order are kept exactly).
• Uncen core from: SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF @ eec4e10a (IQ3_S tier).
• What changed: 144 tensors — ffn_down_exps (48, expert-sliced 256 of 512), ffn_down_shexp (48), ssm_out (36) and attn_output (12). The other 1,079 tensors are byte-identical to the Coder (attention, router, gate/up experts, embeddings, norms, tokenizer, chat template, metadata — untouched).
• Payload: 29.85 GB + 28.80 GB = 58.65 GB (54.62 GiB), 2 shards; PLE table lives in shard 2 (as in the Coder release).
• Expert map: David's 256-expert map, verified to agree with ISTA's per-expert allocation in all 48 layers; the 12,288 expert slices are copied byte-for-byte (no dequantise/re-quantise; GSQ scales preserved where present).
• Verification: per-tensor blake2b hashes plus per-expert provenance for all 12,288 slices are recorded in MANIFEST.json. This artifact was rebuilt on 2026-10-12 from the archived build manifest and verified byte-identical to the original build (sha256 2b3d54ff… for shard 1; shard 2 re-hashed to the upstream value 316b46f3…).
• Runtime status: served and smoke-tested with the Strata engine (0.1.40) on 2026-10-12.

Runtime support. Built for and served with the Strata engine. Stock llama.cpp / Ollama / vLLM are not tested with this artifact.

Community testing & feedback

Behavior reports are welcome. Please open a discussion in the Community tab with: hardware + RAM, engine + version, context/settings, and what you observed (refusals or their absence, behaviour, languages).

---

1. What this is (and is not)

• It is the Coder release with the abliterated tensors grafted in: the model keeps the Coder's code/agent tuning and gains the uncensored character of the SC117 abliterated weights.
• It is not a re-quantisation: every transplanted byte comes from the pinned SC117 revision; the kept expert blocks are sliced exactly, and the 96 small projections are taken as SC117's Q8_0 bytes.
• It is not a fine-tune: no training was run; nothing was distilled.

2. Specification

Files

3. Quick start (Strata)

# 1) weights
hf download DatKurai/Qwen3.8-Flash-Next-GSQ-RCO-Coder-Uncen-IQ1_M-GGUF --local-dir coder-uncen

# 2) native pack for the engine (verified on this release)
python tools/iq_pack.py \
  --gguf coder-uncen/Qwen3.8-Flash-Next-GSQ-RCO-Coder-Uncen-IQ1_M-00001-of-00002.gguf \
  --out  coder-uncen/pack

# 3) serve (PLE is in shard 2; the shipped expert profile enables the resident-expert mode)
strata --serve \
  --pack           coder-uncen/pack \
  --native         coder-uncen/Qwen3.8-Flash-Next-GSQ-RCO-Coder-Uncen-IQ1_M-00001-of-00002.gguf \
  --ple-gguf       coder-uncen/Qwen3.8-Flash-Next-GSQ-RCO-Coder-Uncen-IQ1_M-00002-of-00002.gguf \
  --expert-profile coder-uncen/expert-profile-coder.bin \
  --kv             int8 \
  --resident-experts \
  --max-context    131072

Notes:

• --ple-gguf points at shard 2 here (unlike some other releases in this family, the Coder keeps the PLE table in shard 2).
• --resident-experts requires a static profile; the one shipped in this repo is exactly the ranking this build was served with. Drop both flags to run without it.
• An MTP draft is optional; the standard Qwen3.8-Flash-Next draft layer works with the --mtp flag.
• An advanced setup is required in the sense that Strata itself must be built or downloaded from its repository first; the engine is not part of this repo.

4. Limitations

• Verification scope: byte-level only — per-tensor and per-expert hashes in MANIFEST.json (the graft changes refusal-behaviour tensors; everything else is the Coder's own bytes).
• Uncensored by design: the abliteration removes refusal behavior; outputs are not safety-filtered. Use lawfully and responsibly.
• Engine and context: verified with Strata; other runtimes untested. Context metadata is the Coder's (262,144); serving settings determine what you actually get.
• Vision is not re-tested on this artifact and no projector is shipped.

5. Provenance & credits

Removing refusals removes a safety behaviour. What the model writes is your responsibility.

6. Citation

@misc{datkurai_coder_uncen_2026,
  title        = {Qwen3.8-Flash-Next-GSQ-RCO-Coder-Uncen - IQ1_M - 256 experts/layer (GGUF)},
  author       = {DatKurai},
  year         = {2026},
  howpublished = {Hugging Face},
  note         = {Weights-only abliteration transplant into the ISTA-DASLab GSQ-RCO Coder release}
}