Arioron, a Dhaka-based AI company, has released Vex-Amber-Fable-2.0, a 2-billion-parameter small language model (SLM) claiming a world record in the sub-3B category on SWE-bench Verified with 65.37% accuracy. This positions it as delivering about 85% of frontier models like Claude Sonnet 4.5 (77%) and GPT-5.1 (76%), despite being 50 times smaller.The model uses float32 precision and an 8k context window, matching 8B-30B models on coding benchmarks like HumanEval (60.98%) and LiveCodeBench (44.19%).
Source: @arioron_intl
