Nanbeige4.3-33B Opens Under Apache 2.0 With 512K Context

What shipped

Nanbeige, the LLM lab owned by Chinese recruiting platform BOSS Zhipin, has released Nanbeige4.3-33B under Apache 2.0, with a base checkpoint alongside it. The model card reports 33.1 billion parameters in bfloat16, a Llama-family dense architecture, and a 524,288-token context window. Thinking mode is on by default and can be disabled, and a separate setting controls whether earlier turns keep their reasoning, with the card recommending keeping it for multi-turn tool use and dropping it for ordinary chat.

The unusual part of the positioning is in the card itself: rather than pitching the 33B as an end product, Nanbeige calls it "the flagship teacher model for the Nanbeige4.3 series", meant for "the distillation of advanced agentic capabilities into smaller models". That is a recruiting company building an AI lab, using a large open model chiefly as a source of training signal for the compact models it presumably actually ships. It is also a dense model in a field that has swung hard toward sparse mixtures-of-experts: no expert routing is mentioned anywhere in the card, which keeps the architecture simple and the memory story predictable, at the cost of activating every parameter on every token.

The benchmark claims, with the usual caveats

The card's comparison table puts Nanbeige4.3-33B at 80.9 on Terminal-Bench 2.1, against 73.0 for Qwen3.8-27B, 61.8 for Meta's Muse-Glimmer-30B and 47.2 for Gemma4-31B. Other reported scores include 89.5 on GPQA Diamond, 95.2 on AIME 2026, 65.5 on DeepSWE and 60.3 on SWE-Bench Pro. These are self-reported numbers on the vendor's chosen harness, so they belong in the same bucket as every open-model release this year: strong signal, pending independent replication. The card also reports 23.8 on Terminal-Bench 4.0, a reminder that one benchmark version is not comparable to another.

Running it yourself

The recommended inference settings split by workload, and they are revealing about the intended use. Agentic and tool-driven sessions run at temperature 1.0 with up to 65,536 new tokens, while ordinary reasoning and chat run at temperature 0.6 with a 262,144-token generation budget, and the card recommends XML for tool calls with JSON as the supported alternative. The weights total about 66GB in bfloat16, and the card's serving examples for SGLang and vLLM use tensor parallelism across four GPUs.

For smaller setups, the interesting fact is on the hub rather than in the card: 56 community quantizations already exist, including GGUF builds compatible with llama.cpp, Ollama and LM Studio, so a quantized copy for a single workstation is already available even though the lab ships none.

For readers comparing open agentic models, this lands next to two releases we have covered this month: DeepSeek's DeepInfer serving engine and the export-clearance regime those labs now operate under. Nanbeige4.3-33B is a smaller, denser entry than the 300B-plus MoE models that dominate agentic leaderboards, which makes it notable precisely because a dense 33B at 512K context is the size a serious local setup can actually serve.