Nanbeige's first large model since Nanbeige2-16B, and its first in the agentic era. It is a 33.1B dense model with a 512K-token context, released under Apache 2.0 on October 24, 2026. Base and instruct weights shipped together, and FP8, INT8 and INT4 checkpoints followed on October 31. The architecture is plain full-attention Llama with GQA: 52 layers, hidden size 7168, 56 query heads and 8 KV heads. It does not use the looping of Nanbeige4.2-3B.

The card calls it the "flagship teacher model for the Nanbeige4.3 series". It is meant first to carry agentic coding and "cowork" skills, and then to distill them into a coming Nanbeige4.3-3B.

No technical report has been published, and the cards give neither training tokens nor compute. The released base model and the lab's own tokenizer family (a 166,912-token vocabulary) indicate in-house pretraining. The cards compare the base model only with Gemma4-31B-Base, which it beats on MMLU-Pro (66.9 against 59.3), GPQA (62.0 against 41.5) and CMMLU (86.8 against 76.9).

The instruct model was evaluated at maximum thinking effort with the same harness for every model. Nanbeige's self-reported results lead Qwen3.8-27B, Muse Glimmer and Gemma4-31B on most rows:

  • Terminal-Bench: 80.9 on version 2.1 and 23.8 on 4.0 (Qwen3.8-27B 6.3).
  • DeepSWE: 65.5 against 42.5.
  • AutomationBench: 54.2 against 38.3.
  • AIME 2026: 95.2.

It trails Qwen3.8-27B on SWE-bench Pro (60.3 against 61.7) and GDPval (84.8 against 86.4).

Artificial Analysis scores it 20 on Intelligence Index v5.0.1, level with Qwen3.8-27B (xhigh), and 30% on Terminal-Bench 4.0. Its output is verbose, at 158K tokens per index task. AA lists no public API provider. The instruct repository passed 1,000 HuggingFace likes within two weeks of release.

Model Details

Architecture DENSE
Parameters 33B
Context window 524,288
AA Intelligence 20
License Apache 2.0

Benchmark Scores

Benchmark Score Mode
Terminal-Bench 4.0 23.8 —
SWE-bench Pro 60.3 —
GDPval 84.8 —
GPQA Diamond 89.5 —
AIME 2026 95.2 —
AA-LCR 84.0 —

Variants

Name Parameters Notes
Nanbeige4.3-33B-Base — Base model, 128K context
Nanbeige4.3-33B — Instruct/reasoning model, 512K context; FP8, INT8, INT4 checkpoints (Oct 31)
agentscodingreasoninglong-contextopen-weight

Related