Nanbeige4.3-33B: Specs, License & How to Run

Nanbeige4.3-33B is a dense 33B text model from Nanbeige LLM Lab, the Hugging Face org for BOSS Zhipin’s model team (huggingface.co/Nanbeige, opened 2026-10-26). Official hosts the same day: the ungated card Nanbeige/Nanbeige4.3-33B (created 2026-10-24T02:49:18Z; last modified 2026-10-25T11:11:17Z; 365 likes / 0 downloads) and the Base sibling Nanbeige/Nanbeige4.3-33B-Base (created 2026-10-24T02:46:52Z; 46 likes). Both tag apache-2.0. A raw LICENSE path 404’d on both.

This page hangs off ChatGPT alternatives, closed LLM APIs, Claude Code alternatives, and Cursor alternatives. That mapping is editorial. We have no sourced quality comparison against ChatGPT, Claude Code, or Cursor. Head-to-head facts versus the same-site ~30B card the Nanbeige table names: Nanbeige4.3-33B vs Qwen3.8-27B. Muse-Glimmer-30B and Gemma4-31B are also columns on that table and are not pages here.

The card calls Nanbeige4.3-33B the flagship teacher for the Nanbeige4.3 series: a high-capacity agent to later distill into smaller models (it names Nanbeige4.3-3B as upcoming). It positions “complex coding and professional cowork.” Those sentences stay vendor. Hugging Face pipeline_tag is text-generation; model_type is llama. Languages on the card: en, zh.

What we verified (2026-10-26)

Field	Value	Source
Host	huggingface.co/Nanbeige/Nanbeige4.3-33B (not gated)	HF API
Created / last modified	2026-10-24T02:49:18Z / 2026-10-25T11:11:17Z	same
Likes / downloads	365 / 0	same
License tag + file	apache-2.0 / LICENSE blob 404	HF cardData + raw fetch
Pipeline tag	text-generation	HF API
Architecture	llama / LlamaForCausalLM	config.json
Params (card)	33B	card
safetensors.parameters	33,118,977,024	HF API
usedStorage (BF16)	66,259,989,086 bytes (~61.7 GiB)	same
Layers / hidden / heads	52 / 7168 / 56 Q, 8 KV	config.json
Documented context	524,288 tokens	card + max_position_embeddings
Base sibling	Nanbeige4.3-33B-Base · 46 likes · same param count	HF API
Org	Nanbeige LLM Lab at BOSS Zhipin	huggingface.co/Nanbeige

How to run (from the card, not tested)

The tokenizer chat template exposes enable_thinking (default on), preserve_thinking (the card recommends false for general chat and true for multi-turn tool use, office tasks, and code-agent workflows), and a tools path. The card recommends tool_call_format="xml"; json is also listed. Official benches used preserve_thinking=true.

Recommended sampling (copied):

Scenario	Temperature	Top_p	Max new tokens
Agentic and tool-use	1.0	0.95	65,536
Reasoning and chat	0.6	0.95	262,144

SGLang (card; --tp-size 4):

python -m sglang.launch_server \
  --model-path /path/to/Nanbeige4.3-33B \
  --host 0.0.0.0 \
  --port 8000 \
  --tp-size 4 \
  --mem-fraction-static 0.8 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

vLLM (card; --tensor-parallel-size 4):

vllm serve /path/to/Nanbeige4.3-33B \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 4 \
  --gpu-memory-utilization 0.8 \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

Transformers is also documented (trust_remote_code=True, use_fast=False on the tokenizer). Engine page: vLLM.

License

Card tag apache-2.0 on the 33B and Base cards (2026-10-26). No LICENSE blob in either repo on this check. Not legal advice.

What we did not invent

• “State-of-the-art among ~30B open models” stays vendor.
• Card bench rows stay self-reported. On that table, Qwen3.8-27B leads SWE-Bench Pro and GDPval; Nanbeige leads most other listed rows, including Terminal-Bench 2.1 (80.9 vs 73.0) and DeepSWE (65.5 vs 42.5). We do not merge those into a rank.
• Peak VRAM, tokens/s, and laptop-class GGUF paths are not on this card.
• We do not rank Nanbeige against Cursor, Claude Code, ChatGPT, Flint, or Raon-1-mini.

Related ~30B open weights: Qwen3.8-27B — Nanbeige4.3-33B vs Qwen3.8-27B — Flint (39.8B/2.5B coding MoE), Raon-1-mini (28.8B/4.3B preview), Sage-A1.5 (35B science/agent VLM). Serving: vLLM. Agents: Cline, OpenCursor.