Nanbeige releases a 33B model as community tests a smaller local build

What happened?

Nanbeige’s model card presents Nanbeige4.3-33B as a dense 33-billion-parameter model intended for agentic coding and professional tasks. Its repository was created October 24. The model is offered under Apache 2.0, and its configuration identifies a LlamaForCausalLM architecture with BF16 weights.

The developer publishes comparisons against other models on coding and work-oriented benchmarks. These are vendor-reported results, not independently reproduced evidence of overall superiority.

A separate community quantizer reports running an approximately 10.1 GiB mixed-precision IQ2_XXS GGUF on one RTX 3060 with 12GB of memory, under specified context and low-bit KV-cache settings. The conversion uses selected Q4_K and Q6_K layers alongside other tensor types. That is a community hardware-fit report, not a LessCarbs test or a guarantee that the full BF16 model fits on the card.

Source: Nanbeige ↗ · Source: DogukanUrker ↗

Why does it matter?

Quantization can make larger models accessible on smaller hardware, but it changes the representation and may affect quality. A reported fit for one configuration is not a universal performance or memory claim.