Original Teacher Model by Nanbeige / BOSS Zhipin • GGUF Quantization & Protium-1.0 Cognitive Architecture by Solstice-AI
SolsticeAI/Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium is the official Solstice-AI GGUF release of Nanbeige4.3-33B, the open-weights 33.1B dense teacher model developed by Nanbeige LLM Lab (BOSS Zhipin / Kanzhun Limited).
Engineered specifically for mission-critical software engineering, complex multi-file repo refactoring, terminal automation, and long-horizon cowork, this release integrates Solstice-AI's Protium-1.0 Objective-Gated Dynamic Reasoning Architecture with native high-fidelity GGUF quantizations.
Calibrated with importance matrices to protect delicate coding logic, syntax trees, and tool-call signatures:
| Quant Checkpoint | File Size | Recommended Hardware | Description & Target Profile |
|---|---|---|---|
nanbeige4.3-33b-q8_0.gguf |
32.78 GB | 48GB+ VRAM or 64GB+ System RAM | Full Reference Master: Absolute highest precision reference tier. Zero perceptible quality degradation vs. unquantized BF16. Best for dedicated inference servers. |
nanbeige4.3-33b-q6_k.gguf |
25.31 GB | 32GB - 48GB VRAM or 64GB Unified RAM | Near-Lossless Tier: Retains deep multi-step mathematical proofs and abstract syntax trees. Ideal for Apple Silicon Mac (64GB/128GB) and multi-GPU setups. |
nanbeige4.3-33b-q5_k_m.gguf |
21.89 GB | 24GB - 32GB VRAM or 32GB RAM | High-Efficiency Sweet Spot: Delivers >99% teacher performance with noticeably lower memory bandwidth pressure and faster tokens-per-second. |
nanbeige4.3-33b-q4_k_m.gguf |
18.67 GB | Single 24GB GPU (RTX 3090 / 4090) | ⭐️ Recommended Production Driver: Fits cleanly into a single 24GB consumer GPU with working KV context headroom. Exceptional balance of speed and reasoning depth. |
[!WARNING]
Technical Advisory Regarding Q3_K_M
For extreme memory-constrained environments,
Q3_K_M(15.2 GB) is hosted in a separate standalone repository:SolsticeAI/Nanbeige4.3-33B-Q3_K_M-GGUF.However,
Q3_K_Mis NOT recommended for agentic software engineering or complex mathematical reasoning. Quantizing dense 33B teacher models below 4 bits introduces measurable degradation in multi-turn tool calling and syntax generation. For serious coding workloads,Q4_K_Mis the recommended minimum operational baseline.
This release incorporates Protium-1.0, Solstice-AI's high-efficiency cognitive protocol:
<think> and </think> blocks prevents conversational runaway and circular deductions.llama.cpp (llama-server)# Recommended baseline: Q4_K_M on 24GB GPU
llama-server \
--model 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
Create a Modelfile:
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 "<|endoftext|>"
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{- end }}
{{- range .Messages }}
<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{- end }}
<|im_start|>assistant
<think>
"""
Build and run:
ollama create nanbeige4.3-33b -f Modelfile
ollama run nanbeige4.3-33b
Nanbeige4.3-33B benchmarked under maximum reasoning effort against the frontier ~30B class:
| Category | Benchmark | Nanbeige4.3-33B | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B |
|---|---|---|---|---|---|
| Agentic Coding | Terminal-Bench 2.1 | 80.9 | 73.0 | 61.8 | 47.2 |
| Terminal-Bench 4.0 | 23.8 | 6.3 | 1.6 | 0.0 | |
| DeepSWE | 65.5 | 42.5 | 4.4 | 2.7 | |
| SWE-Bench Pro | 60.3 | 61.7 | 36.9 | 51.2 | |
| SWE-Atlas-QnA | 50.8 | 41.9 | 17.7 | 3.3 | |
| NL2Repo | 53.0 | 48.3 | 27.4 | 33.8 | |
| Agentic Cowork | Automation Bench | 54.2 | 38.3 | 15.2 | 9.7 |
| JobBench | 44.8 | 34.5 | 16.6 | 9.0 | |
| GDPval | 84.8 | 86.4 | 75.8 | 71.4 | |
| Claw-Eval (pass^3) | 78.3 | 75.2 | 57.3 | 44.2 | |
| Toolathlon-Verified | 70.4 | 68.5 | 47.2 | 27.8 | |
| DeepPlanning | 63.3 | 49.2 | 21.4 | 19.4 | |
| Reasoning | GPQA Diamond | 89.5 | 89.2 | 83.5 | 91.3 |
| AIME 2026 | 95.2 | 94.5 | 94.7 | 89.2 | |
| AA-LCR | 84.0 | 82.0 | 83.3 | 72.0 |
This model card and quantization suite inherit the Apache 2.0 License from upstream Nanbeige. Free for commercial and research use.
@misc{nanbeige43ggufprotium2026,
title = {Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium},
author = {Solstice-AI Team and Nanbeige Team},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/SolsticeAI/Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium}}
}
Solstice-AI • Sovereign High-Efficiency Machine Intelligence
Nanbeige/Nanbeige4.3-33B