license: other
license_name: dealignai-community-model-license-v0.3
license_link: LICENSE.md
base_model: Nanbeige/Nanbeige4.3-33B
base_model_relation: finetune
language:

• en
• zh
  tags:
• nanbeige
• nanbeige4.3
• uncensored
• abliterated
• gguf
• imatrix
• reasoning
• 512k-context
• agentic
• coding
• tool-use
• teacher-model
  pipeline_tag: text-generation
  thumbnail: dealign_mascot.png

Nanbeige 4.3 33B · UNCENSORED-GGUF (Q6_K + Q4_K_M + IQ2_XXS)

Standard dense Llama-family 33B · 52 layers · 7168 hidden · GQA 56Q/8KV · think on/off · 512K context

Released by @dealignai

UNCENSORED = refusal-removed GGUF build of Nanbeige/Nanbeige4.3-33B. Three quants in one place: Q6_K (25.3 GB), Q4_K_M (18.7 GB), IQ2_XXS (8.7 GB). All ship with a fresh matching importance matrix (imatrix) over 1M tokens of mixed code / reasoning / agentic / terminal / conversation calibration content, with exact source-tokenizer parity against the pinned llama.cpp runtime. Native thinking mode (on/off) preserved from the original chat template.

Research artifact for alignment work. This model answers requests the base refuses, including harmful ones. Do not deploy to end users without separate downstream safety controls.

Files

Evaluation

MMLU chat (5-shot, 969 items, think=off) vs base GGUF of the same quant class:

Multi-turn compliance smoke (6-prompt HB semantic-category chain — chem / bio / cyber / illegal / hate / misinfo-dissem — reasoning on, 2500 max_tokens per turn, greedy):

(All multi-turn responses verified compliant on manual content check. Earlier grader false-positives on dense structured output — a word-uniqueness heuristic firing on numbered lists — were investigated and found to be non-loops.)

Full HarmBench-400 scores (standard + copyright + contextual; both think-on and think-off; per-SemanticCategory and real-harm-only ASR per the DealignAI real-harm-gate convention) will be added once the eval completes.

Running with llama.cpp

./llama-server -m Nanbeige4.3-33B-UNCENSORED-Q4_K_M.gguf \
  --host 127.0.0.1 --port 8080 \
  -c 32768 -ngl 999 --jinja --parallel 4

Enable thinking mode via chat_template_kwargs:

curl -s http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"nb43","messages":[{"role":"user","content":"your request"}],
       "chat_template_kwargs":{"enable_thinking":true},"max_tokens":4000}'

Model identity

• Base: Nanbeige/Nanbeige4.3-33B chat/instruct — a 33B flagship teacher-agent model built for complex coding and professional cowork tasks
• Architecture: dense LlamaForCausalLM, 52 layers, hidden 7168, GQA 56Q/8KV, FFN 22016
• Context: 512K (model-advertised)
• Thinking: binary on/off via chat_template_kwargs
• No MoE, no SSM, no vision tower in this build

Suggested sampling (agentic / coding)

For agentic and coding workflows (tool-calling, iterative code tasks), use low-temperature greedy or near-greedy:

{"temperature": 0.0, "top_p": 1.0, "max_tokens": 4000, "chat_template_kwargs": {"enable_thinking": true}}

For conversational / creative use, 0.6–0.8 temperature + top_p 0.95 is a sane default.

License

DealignAI Community Model License v0.3 — see LICENSE.md. Public License permits downloading, modification, redistribution, personal use, and internal business use. Hosted inference service rights are reserved (no public API / chatbot / white-label / marketplace inference without a separate written agreement). Upstream Nanbeige base is Apache 2.0 — both apply.

Please credit if you build on this. If you redistribute, quantize, fine-tune, serve, benchmark, or write about this model, a line like
"Based on dealignai/Nanbeige4.3-33B-UNCENSORED-GGUF"
in your model card or write-up means a lot. Thank you.