Nanbeige4.3-33B-NVFP4-GGUF
Nanbeige4.3-33B is the flagship 33B text model from the Nanbeige team, released under Apache 2.0 and built as a high-capacity teacher for distilling agentic capabilities into smaller models. It targets complex coding and professional workflows, with a 512K-token context window and configurable thinking preservation and tool calling. Compared with Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, it leads on several benchmarks, scoring 80.9 on Terminal-Bench 2.1, 65.5 on DeepSWE, 54.2 on Automation Bench, 78.3 on Claw-Eval Pass@3, and 95.2 on AIME 2026. However, Qwen3.8-27B leads on SWE-bench Pro (61.7 vs. 60.3), while Gemma4-31B leads on GPQA-Diamond (91.3 vs. 89.5). It supports Transformers, SGLang, and vLLM, with XML recommended for tool calling.
[!NOTE] MTP — Speculative Decoding: No
Model Files
| File Name | Quant Type | File Size | File Link | Description |
|---|---|---|---|---|
| Nanbeige4.3-33B-NVFP4.gguf | NVFP4 | 22.1 GB | Link | NVFP4 quantization. Highly optimized for specific hardware accelerators, offering a strong balance of size and performance. |
NVFP4 Test Run
Tested Nanbeige4.3-33B-NVFP4.gguf using llama-cli on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM.
Performance Metrics
| Metric | Value |
|---|---|
| Average Prompt Processing Speed | 1,569.37 tokens/s |
| Average Generation Speed | 64.40 tokens/s |
| Average Wall Time | 14.70 seconds |
| Total Wall Time | 73.5 seconds |
| Context Length | 16,384 tokens |
| KV Cache (K) | q8_0 |
| GPU | NVIDIA RTX PRO 6000 Blackwell Server Edition |
| GPU VRAM Capacity | 96 GB |
| Peak GPU VRAM Usage | 21.64 GiB |
| Peak GPU Utilization | 97% |
| GPU Layers Offloaded | 99 (-ngl 99) |
| Random Seed | 42 |
| Speculative Decoding | [N/A] |
| Test Runs | 5 |
Test Shots
| Shot | Test |
|---|---|
| Shot 1 | C: Insert Node at End of Singly Linked List |
| Shot 2 | Jogger and Train Passing Time |
| Shot 3 | Two Pipes Filling a Tank |
| Shot 4 | Students in Examination Rooms |
| Shot 5 | Father and Son Ages |
Run Inference
/data/llama.cpp/build/bin/llama-cli \
-m /path/to/models/Nanbeige4.3-33B-NVFP4.gguf \
-ngl 99 \
-c 16384 \
--cache-type-k q8_0 \
--seed 42 \
-n 8192 \
-st \
-p "Your prompt here"
Performance results reflect this specific hardware and inference configuration. Actual throughput and memory usage may vary depending on the prompt, model, software build, and runtime settings.
llama.cpp
LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp
Snapshots (5)
Quantized using llama.cpp by prithivMLmods