Nanbeige LLM Lab has released Nanbeige4.3-33B, a 33-billion-parameter open model built for agentic coding and office-style “cowork” tasks. The dense model is available on Hugging Face under the Apache 2.0 license, alongside a base version for custom post-training.
Nanbeige4.3-33B (Credits: MarkTechPost)
Nanbeige4.3-33B open agentic coding model specs
The model is a dense architecture — all 33 billion parameters run on every token, with no mixture-of-experts routing — and it supports a massive 524,288-token (512K) context window. It uses grouped-query attention with 56 query heads and 8 key-value heads, and its chat template includes toggles for reasoning and multi-turn tool use, with an XML call format recommended for code agents.
Practically, this is a multi-GPU server model: the published SGLang and vLLM recipes use 4-way tensor parallelism, and the BF16 weights need roughly 66 GB. It is aimed at developers and teams running their own agentic coding infrastructure, not at consumer laptops.
Nanbeige4.3-33B benchmark results — with a caveat
In Nanbeige’s own comparison table, the model leads 17 of 20 benchmarks against three open models near 30 billion parameters: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Headline results include 80.9 on Terminal-Bench 2.1 (versus 73.0 for Qwen3.8-27B) and a wide gap on DeepSWE, 65.5 versus 42.5. On agentic “cowork” tests it posts 54.2 on Automation Bench and 44.8 on JobBench.
It does not win everywhere: it trails Qwen3.8-27B on SWE-Bench Pro (60.3 vs 61.7) and GDPval, and Gemma4-31B leads GPQA Diamond in the same table. All of these figures are self-reported by Nanbeige, and the report itself flags discrepancies — for example, Meta’s own card puts Muse-Glimmer-30B at 51.2 on SWE-Bench Pro, not the figure in Nanbeige’s table. Independent testing will be the real verdict.
Nanbeige4.3-33B is positioned as the flagship “teacher” model of the series: the lab plans to distill its agentic skills into smaller models, starting with Nanbeige4.3-3B. If that distillation works, the practical payoff could be a far smaller model that still codes and works across tools on its own.
Also Read: Anthropic Launches OSS Scanner, a Free AI Vulnerability Scanner for Open Source | I Stopped Uploading Private Documents to AI — This Free App Handles Them on My PC With Ollama
Source: MarkTechPost — read the full story there.
