Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw · Hugging Face

Source: https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw

https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw#glm-53-uncensored-exl3-30bpwGLM-5.3-UNCENSORED EXL3 3.0bpw

EXL3 quantization ofdealignai/GLM-5.3-UNCENSORED-FP8, itself a weight-edited (no fine-tune) variant ofzai-org/GLM-5.3. The edit is documented inCRACK\_SURGERY\.json(copied unchanged from the source repo). This repo is not affiliated with dealignai or Z.ai.

• Architecture:GlmMoeDsaForCausalLM, 753B total parameters, 256 routed experts (8 active) + 1 shared, MLA attention with DSA sparse indexer, 78 layers + 1 MTP layer
• Average bitrate: 3.04 bpw (\-b 3\.0 \-\-hq; attention and shared experts at 5 bpw, dense MLPs at 4, routed experts at 3), lm_head 6 bpw,mul1codebook
• MTP (next-token prediction) layer included (experts 4 bpw, attention and shared expert 6 bpw, uncalibrated), usable as a speculative draft (draft\_mode: mtpin TabbyAPI)
• Size: 273 GiB
• Converted with ExLlamaV3 at commitd3739fd, default calibration (250 rows x 2048 tokens), source read directly from the FP8 checkpoint

https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw#fidelity-vs-the-fp8-sourceFidelity vs the FP8 source

eval/model\_diff\.py, 20 rows x 2048 tokens of wikitext-2 test:

metricvalueKL divergence (quant ‖ FP8)0.089KL divergence (FP8 ‖ quant)0.097per-token KL, median / p900.021 / 0.221perplexity, quant / FP83.440 / 3.302median KL where FP8 top-prob ≥ 0.95 (44% of tokens)0.0011

https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw#agentic-evaluationAgentic evaluation

tau2-bench (airline, retail), Pass^1. Agent temperature 1.0, top_p 0.95; user simulator and judges GPT-4.1 at temperature 0. The reference is stock GLM-5.3 (not the uncensored edit) served at FP8 by Z.ai via OpenRouter, run through the same harness. The difference therefore mixes the dealign weight edit and this quantization.

domainstock GLM-5.3 FP8 (Z.ai)this quantdifferenceairline (50 tasks x 2 trials)0.710 ± 0.0450.640 ± 0.048-0.070 (~1.1 SE)retail (114 tasks)0.504 ± 0.033 (2 trials)0.482 ± 0.047 (1 trial)-0.022 (~0.4 SE) Per trial: airline baseline 0.740 / 0.680, quant 0.600 / 0.680; retail baseline 0.465 / 0.544, quant 0.482. ± is one binomial standard error. Neither difference is statistically significant at these sample sizes; treat the airline point estimate as a possible small regression rather than a measured one.

https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw#servingServing

Tested with TabbyAPI on 8x A100 40GB, layer split (gpu\_split\_auto), MTP drafting, 98K-token shared cache.

model:
  model_name: GLM-5.3-UNCENSORED-EXL3-3.0bpw
  backend: exllamav3
  max_seq_len: 65536
  cache_size: 98304
  gpu_split_auto: true
  tool_format: glm4_7
  reasoning: true
draft_model:
  draft_mode: mtp

**Tool calling caveat:**GLM writes tool arguments as raw text (<arg\_value\>9523456873</arg\_value\>). TabbyAPI’sglm4\_5parser (as of commitbe74bf0) JSON-decodes every value without consulting the tool schema, so string parameters that look like numbers (order and product IDs, zip codes) reach your tools as integers. On tau2-bench retail this caused ~70% of tool calls to fail. Make the parser keep parameters whose schema type isstringas raw text before using this model for agents. The evaluation above was run with that fix applied.

https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw#licenseLicense

MIT, following the upstream GLM-5.3 and dealignai releases.