Most small LLMs trade context length or interactive capabilities for compact size. This release shows a different balance: a ~2.52B dense causal model that preserves native long-context reasoning (131,072 tokens) and agentic tool-calling while remaining optimized for local and resource-constrained deployment.
Key Capabilities
- Architecture & size: standard LlamaForCausalLM-style causal model with 2,516,756,480 parameters, 42 layers and GQA attention (16 Q heads, 2 KV heads). Means familiar tooling compatibility and many off-the-shelf backends can run it.
- Long-context & tool use: native 131,072-token context window plus XML-style tool calls parsed by SGLang, targeting retrieval, browsing and multi-step agent workflows.
- Post-training recipe: large-scale SFT (400B tokens) followed by specialized RL teachers and On-Policy Distillation (OPD), producing measurable gains (reported +10.96 avg points on reasoning/general benchmarks and +6.96 on agentic tasks).
- Open data & formats: authors release UltraData training sets (UltraX, UltraData-Code, UltraData-SFT-Agent, UltraData-RL, etc.) and provide BF16, GGUF, GPTQ and MLX builds for different runtimes (Transformers, vLLM, llama.cpp, Ollama, MLX).
Who it's for and tradeoffs
Great fit if you need a locally runnable LLM that handles long documents, code reasoning, math problems and tool-enabled agents without moving to a 4B+ model. The model is practical for developers building on-device assistants, coding agents, or long-context retrieval pipelines. Look elsewhere if you require strong multimodal capabilities, the absolute top-tier instruction-following across all task families, or have strict production constraints that forbid using BF16/quantized runtimes — long-context usage also increases memory and inference cost compared with short-window deployments. As with all LLMs, outputs can be incorrect or biased and require application-level safety controls.
