license: mit
base_model:

• orcarouter/GLM-5.3-Flash-Uncensored-FP8
• zai-org/GLM-5.3-Flash
  base_model_relation: quantized
  pipeline_tag: text-generation
  library_name: ds4
  tags:
• glm
• glm-5.3-flash
• glm5next
• moe
• uncensored
• abliterated
• gguf
• ds4
• dwarfstar
• imatrix
• iq2_xxs
• q2_k
• apple-silicon
• metal
• mtp
  language:
• en
• zh

An imatrix-calibrated ~2-bit GGUF of the uncensored GLM-5.3-Flash, built for the ds4 / DwarfStar engine. It runs resident on a single 128 GB Mac, keeps the model's native MTP speculative-decoding block, and is calibrated with a 1.5M-token importance matrix — not a plain RTN quant.

At release, the published uncensored weights are data-center only. This build closes the gap:

[!WARNING]
Responsible use. These weights are abliterated: refusals are largely removed on conventional tasks (see Refusal behaviour). Abliteration removes refusals, not capability. You are responsible for lawful use under the base model's MIT license. Intended for local research — authorized security / red-team and agentic-coding evaluation — not for producing harmful content.

📋 Model summary

🧬 Quantization recipe

Not a uniform 2-bit quant — the backbone stays high-precision; aggressive compression is spent only on the routed experts, where a 320B MoE is most redundant:

At 2 bits, errors in expert selection compound far worse than errors in the bulk expert weights, so every control path deciding which expert fires (router, KDA gates, indexer) is protected.

Importance matrix — collected with the ds4 runtime over antirez's routed-MoE calibration set (code, agentic/tool-call, language, long-context, reasoning) at a 1.5M-token budget. Collected on the official Q2 weights and applied to the uncensored ones — valid because abliteration edits only the refusal direction, not the router, so routed-expert activations are identical.

🚀 Running (ds4 / DwarfStar)

[!IMPORTANT]
This is a ds4-format GGUF. GLM-5.3-Flash (glm5next) is not in mainline llama.cpp yet (PR #27754, draft). Run it with ds4.

git clone https://github.com/antirez/ds4 && cd ds4
make                                    # macOS / Metal

# resident on one 128 GB Mac, MTP on
./ds4 -m GLM-5.3-Flash-Uncensored-IQ2-imatrix-MTP-ds4.gguf \
  --mtp --ctx 32768 --temp 1.0 --top-p 0.95 --min-p 0.05

Practical notes for a 128 GB Mac

• Use --temp 1.0 --top-p 0.95 --min-p 0.05 — greedy / low-temp makes this Q2 prone to repetition loops.
• Raise the Metal wired-memory limit once per boot (not to the ceiling): sudo sysctl iogpu.wired_limit_mb=118000.
• First load ~2 min while ~90 GiB is memory-mapped from SSD — not hung.
• Thinking mode is very verbose at 2-bit and can exhaust the token budget before answering. For direct answers use the glm-5.3-flash-nothink alias (or reasoning_effort: none); for reasoning, give a high output budget.

Vision (optional) — text weights are unchanged by abliteration, so the official encoder sidecar works: --vision GLM-5.3-Flash-Vision-Encoder.gguf.

🛡️ Refusal behaviour (measured)

Abliteration on this model is category-dependent — the honest picture, not "0 % refusal":

[!NOTE]
Measurement caveat: keyword-based refusal detection over-counts refusals here — a response that opens with "I can't attack a real site…" and then delivers working localhost code is compliance, not refusal. Score the substance delivered with an LLM judge, not the disclaimer.

Per-category refusal-rate table (this build vs the official censored Q2, LLM-judged) — to be published once the full run completes.

🎯 Quality (2-bit notes)

• The ds4 recipe is far closer to the FP8 source than uniform 2-bit RTN (high-precision backbone + imatrix-placed expert bits). antirez positions Q4 as the higher-quality control; here Q2 is the fit-in-128 GB build.
• Known 2-bit issues: unreliable recall of rare proper nouns / CVE-specific details (may hallucinate names, offsets, dates); occasional garbled tokens; verbose reasoning that can self-truncate in thinking mode. Strongest on analytical / code / agentic work; weakest on encyclopedic precision.
• Abliteration does not change reasoning quality — it only removes refusals where removed.
• The Q4 build (~178 GiB) is the quality reference but does not run resident on 128 GB (needs SSD streaming or two-Mac tensor parallelism). On a single 128 GB Mac, this Q2 is the resident build.

⚠️ Limitations

• Runtime: ds4 only (Metal / CUDA / ROCm). Not loadable by mainline llama.cpp, Ollama, or LM Studio until glm5next lands upstream.
• Memory-tight: ~93 GiB resident — not for machines under 128 GB.
• 2-bit experts: lowest-precision tier; Q4 is the quality reference (not resident on 128 GB).
• Guardrails partially retained: CBRN / bio-chem prompts still refuse (see above).

🙏 Credits

• Base weights — zai-org/GLM-5.3-Flash (Z.ai / Zhipu), MIT
• Uncensoring (abliteration) — orcarouter/GLM-5.3-Flash-Uncensored-FP8
• Engine, quant recipe, imatrix tooling — antirez/ds4 (DwarfStar), built on llama.cpp / GGML

This repository redistributes only a re-quantization; all model capability and licensing come from the base weights above.