Nanbeige4.3-33B: an open, Apache-2.0 512K coder for agent work
The 33B teacher model tops Qwen3.8-27B on Terminal-Bench and DeepSWE and distills into smaller coders — plus SparkLLM's sub-4B on-device agents.
Copy markdown
A frontier-class open coder you can self-host
Nanbeige4.3-33B is Apache-2.0, with a 524K-token context and reasoning on by default. Vendor numbers put it at 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE, versus 73.0 and 42.5 for Qwen3.8-27B. The catch: it wants tensor-parallel 4 on vLLM/SGLang, so it's a multi-GPU server model, not a laptop one.
On-device agents that fit in 4B
SparkLLM's Spark-X2.6 (1.7B and 4B, Apache-2.0) ships base, instruct, GGUF, INT4/8 and FP8 builds, with up to ~1M-token context and three thinking levels on the 4B. The lab claims the 4B matches ~9B models — pull it into Ollama, llama.cpp or LM Studio and benchmark it against Qwen3.5-4B on your own tool-use tasks.