Nanbeige4.3-33B-GGUF-MLX-GSQ-RCO-EXL3-UNCENSORED-UltraOptimised-Curated-Protium

language:

• en
• zh license: apache-2.0 base_model: Nanbeige/Nanbeige4.3-33B tags:
• solstice-ai
• nanbeige
• nanbeige4.3
• nanbeige4.3-33b
• boss-zhipin
• kanzhun
• uncensored
• abliterated
• protium
• protium-1.0
• cognitive-architecture
• dynamic-reasoning
• objective-gated
• zero-waste
• agentic
• agentic-coding
• cowork
• swe-bench
• swe-bench-pro
• terminal-bench
• gguf
• mlx
• mlx-my-repo
• gsq-rco
• exl3
• exllamav3
• llama.cpp
• ollama
• vllm
• sglang
• q8_0
• q6_k
• q5_k_m
• q4_k_m
• 512k-context pipeline_tag: text-generation

Executive Summary

SolsticeAI/Nanbeige4.3-33B-GGUF-MLX-GSQ-RCO-EXL3-UNCENSORED-UltraOptimised-Curated-Protium is the definitive, all-in-one omnibus release of BOSS Zhipin's Nanbeige4.3-33B (Nanbeige/Nanbeige4.3-33B) — the premier ~30B teacher model built to anchor state-of-the-art agentic software engineering, complex systems programming, and multi-turn workplace cowork.

This repository consolidates the entire hardware deployment spectrum into a single unified hub:

• Curated Standard GGUFs for high-efficiency llama.cpp and ollama inference (Q4_K_M to Q8_0).
• Uncensored GGUFs with surgically abliterated refusal circuits for boundary testing and unrestricted security analysis.
• ExLlamaV3 (EXL3) rungs from 2.5 bpw to 4.5 bpw (SAGE) for blistering GPU token throughput.
• Apple Silicon MLX Suite (3-bit to 8-bit) optimized natively for unified memory Apple M-series chips.
• Grouped Scale Quantization (GSQ-RCO) tiers with Recursive Channel Optimization (IQ2_XXS through Q5_K_M).
• Protium-1.0 Objective-Gated Cognitive Architecture (Solstice-AI/Protium-1.0) — zero wasted tokens, token 1 target anchoring, anti-pivot forward deduction, and dynamic impedance scaling ([L10] Proteus default).

Checkpoint Catalog & Memory Budgets

1. Standard GGUF Suite (Root Directory)

Calibrated with precision importance matrices to protect delicate coding and syntax channels:

2. Uncensored GGUF Suite (Root Directory)

Dealigned weights with refusal vectors surgically neutralized. Zero preachy moralizing or refusal filters:

3. ExLlamaV3 (EXL3) Suite (EXL3/)

State-of-the-art EXL3 format for ultra-fast GPU inference with ExLlamaV3:

4. Apple Silicon MLX Suite (MLX/)

Native Apple MLX format with complete tokenizer, configs, and safetensors shards:

5. Grouped Scale Quantization Suite (GSQ-RCO/)

Converted by abenzerps using Recursive Channel Optimization:

• Nanbeige4.3-33B-GSQ-RCO-Q5_K_M.gguf
• Nanbeige4.3-33B-GSQ-RCO-IQ4_XS.gguf
• Nanbeige4.3-33B-GSQ-RCO-IQ3_M.gguf
• Nanbeige4.3-33B-GSQ-RCO-IQ3_XS.gguf
• Nanbeige4.3-33B-GSQ-RCO-IQ2_M.gguf
• Nanbeige4.3-33B-GSQ-RCO-IQ2_XS.gguf
• Nanbeige4.3-33B-GSQ-RCO-IQ2_XXS.gguf

Protium-1.0 Objective-Gated Cognitive Architecture

This release natively integrates Solstice-AI's Protium-1.0 Cognitive Architecture (see Solstice-AI/Protium-1.0).

Named after Protium ($^1\text{H}$) — the fundamental hydrogen isotope with zero neutrons and zero dead weight — Protium-1.0 replaces traditional, unconstrained chain-of-thought bloat with an objective-driven, forward-only reasoning pipeline.

Core Architectural Principles

• Token 1 Target Anchoring: At step zero, the model isolates and locks the exact mathematical, programmatic, or empirical criteria required for complete verification before generating arbitrary reasoning steps.
• Forward-Only Deduction: Eliminates conversational stalling, rhetorical fillers, and internal monologue ("let's see", "now I wonder", "let me rethink"). Deductive progression remains unbroken.
• Anti-Pivot Invariant: Strictly forbids false hesitation pivots ("Wait...", "Hold on...", "Actually...") unless a concrete, isolated contradiction has been mathematically or empirically uncovered.
• Objective-Gated Early Exit: The moment verification criteria are confirmed, reasoning terminates cleanly and the final response is delivered with high information density.
• Preserve Thinking in Multi-Turn: When running agentic coding and multi-step tool loops, set preserve_thinking=True to carry forward analytical deductions across tool calls.

Cognitive Spectrum & Invocation Markers

Protium-1.0 defaults to [L10] Proteus (Dynamic Adaptive Engine), automatically scaling cognitive depth to query complexity. You can explicitly target any calibrated tier by prefixing your prompt:

Example Invocation:

> [L4] Architect a thread-safe distributed task queue in Rust with Tokio, priority scheduling, and zero lock contention on hot workers.

Quickstart & Serving Guide

1. Download via hf CLI

# Recommended GGUF Baseline (Q4_K_M)
hf download SolsticeAI/Nanbeige4.3-33B-GGUF-MLX-GSQ-RCO-EXL3-UNCENSORED-UltraOptimised-Curated-Protium \
  --include "nanbeige4.3-33b-q4_k_m.gguf" \
  --local-dir ./nanbeige4.3-33b-gguf

# Recommended EXL3 Suite (4.0 BPW)
hf download SolsticeAI/Nanbeige4.3-33B-GGUF-MLX-GSQ-RCO-EXL3-UNCENSORED-UltraOptimised-Curated-Protium \
  --include "EXL3/4BPW/*" \
  --local-dir ./nanbeige4.3-33b-exl3-4bpw

# Recommended Apple MLX Suite (4-bit)
hf download SolsticeAI/Nanbeige4.3-33B-GGUF-MLX-GSQ-RCO-EXL3-UNCENSORED-UltraOptimised-Curated-Protium \
  --include "MLX/4-bit/*" \
  --local-dir ./nanbeige4.3-33b-mlx-4bit

2. Run GGUF with llama.cpp (llama-server)

llama-server \
  --model ./nanbeige4.3-33b-gguf/nanbeige4.3-33b-q4_k_m.gguf \
  --host 0.0.0.0 --port 8080 \
  --ctx-size 65536 \
  -ngl 99 \
  --temp 0.6 \
  --top-p 0.95

3. Run with Ollama

FROM ./nanbeige4.3-33b-q4_k_m.gguf

PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER num_ctx 65536
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
<think>
{{ .Response }}<|im_end|>"""

ollama create nanbeige4.3-33b-protium -f Modelfile
ollama run nanbeige4.3-33b-protium

4. Run with Apple MLX (mlx-lm)

from mlx_lm import load, generate

model, tokenizer = load("./nanbeige4.3-33b-mlx-4bit/MLX/4-bit")

prompt = "> [L4] Implement an optimal concurrent LRU cache in Python using multiprocessing.shared_memory."
messages = [{"role": "user", "content": prompt}]
formatted_prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=formatted_prompt, max_tokens=4096, verbose=True)
print(response)

Or serve via MLX HTTP API:

python -m mlx_lm.server --model ./nanbeige4.3-33b-mlx-4bit/MLX/4-bit --port 8080

5. Run EXL3 with ExLlamaV3 / TabbyAPI

Point your TabbyAPI or ExLlamaV3 runner directly to the EXL3/4BPW directory:

model_name: Nanbeige4.3-33B-EXL3-4BPW
model_dir: ./nanbeige4.3-33b-exl3-4bpw/EXL3/4BPW
max_seq_len: 65536
gpu_split_auto: true

Model Architecture & Empirical Performance

Architectural Specifications

Comparative Benchmark Matrix (~30B Parameter Frontier)

All models evaluated at maximum thinking effort using standardized evaluation harnesses:

License & Responsible Use

This model quantization and associated weights are released under the Apache 2.0 License, preserving the upstream licensing established by BOSS Zhipin / Nanbeige Team.

Citations & Acknowledgments

• Nanbeige Team at BOSS Zhipin (Kanzhun) for developing and releasing the state-of-the-art Nanbeige4.3-33B teacher model.
• ExLlamaV3 Contributors (turboderp, nonkronk, acrogenesis, lxe, ProCreations) for the high-performance EXL3 conversion suite.
• abenzerps for the GSQ-RCO quantizations.
• Apple MLX Team for the MLX framework.
• Solstice-AI for the Omni-Quant curation, integration, and the Protium-1.0 Objective-Gated Cognitive Architecture.

@misc{nanbeige43chat2026,
  title        = {Nanbeige4.3-33B},
  author       = {Nanbeige Team},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanbeige/Nanbeige4.3-33B}}
}

@misc{solstice2026protium,
  title        = {Protium-1.0: Objective-Gated Cognitive Architecture},
  author       = {Solstice-AI},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Solstice-AI/Protium-1.0}}
}