Ripple-65B vs Nanbeige4.3-33B: which new agent model?

Ripple-65B Preview is a dense 65B model with a 1M context under Apache 2.0; Nanbeige4.3-33B is a 33B model with a 512K context and a 23 GB INT4 build.

  • Updated

Verdict. Choose by hardware first, then by license paperwork. Ripple-65B Preview has 64.8B parameters, a 1M-token context and a LICENSE file with Apache 2.0 terms, but its BF16 weights are about 130 GB, so by our estimate you need about 45 GB at 4-bit, or its experimental 57 GB NVFP4 build. Nanbeige4.3-33B has 33.1B parameters, a 512K context and an INT4 build of 22.6 GB made with quantization-aware training, which by our estimate fits a 32 GB GPU, but its repositories carry an Apache 2.0 tag without a LICENSE file. Each reports about 80 on Terminal-Bench 2.1 in its own table, 80.5 for Ripple and 80.9 for Nanbeige, and neither table includes the other model, so the numbers do not show which is better. Ripple's table puts Reflection Beam ahead on SWE-bench Verified, and Nanbeige's table puts Qwen3.8-27B ahead on SWE-Bench Pro.

Specifications

Specifications Ripple-65B PreviewNanbeige4.3-33B
DeveloperRicursive IntelligenceNanbeige LLM Lab
ReleasedNovember 4, 2026October 24, 2026
Total parameters64.8B (Ripple-65B Preview)33.1B (Nanbeige4.3-33B)
Active parametersAllAll
ArchitectureDenseDense
Context window1M tokens512K tokens
InputTextText
License Apache 2.0 (Permissive) Apache 2.0 (Permissive)
Memory for the smallest variant (4-bit)45 GB (Ripple-65B Preview)23 GB (Nanbeige4.3-33B)

Performance claims come from the developer and have not been independently verified.

Pick Ripple-65B Preview if

  • You want a 1M-token context for long agent sessions, which is the developer's figure; Nanbeige's is 512K.
  • You want a LICENSE file in the repository that spells out the Apache 2.0 terms.
  • You have about 48 GB of GPU memory or a 64 GB unified-memory machine for a 4-bit build, or a 128 GB desktop such as the NVIDIA DGX Spark for the 57 GB NVFP4 build.
  • You want a published speculative-decoding draft model (DFlash2) to speed up serving.

Pick Nanbeige4.3-33B if

  • You want the smaller model: the INT4 build is 22.6 GB and fits one 32 GB GPU by our estimate, against about 45 GB at 4-bit for Ripple.
  • You want a choice of official builds: BF16, FP8, INT8, INT4 and a base model.
  • You also need office-style tasks such as documents and spreadsheets, which the lab positions the model for.
  • You want a model that loads in stock Transformers, vLLM and SGLang because it uses the Llama architecture.

Ripple-65B Preview · Nanbeige4.3-33B