OmniCoder-9B v2 drops with agentic stability
Tesslate releases OmniCoder-9B v2, a ground-up rebuild of the 9B coding model optimized for agentic reasoning and long-context stability. The model addresses core repetition issues from v1 while achieving frontier-class performance on reasoning and terminal-based benchmarks.
OmniCoder-9B v2's shift to assistant-only training and a constant learning rate schedule marks a significant leap in model reliability for autonomous coding loops. The version achieves near-zero repetition loops in multi-turn tool calling tasks and inherits a 262K context window from Qwen3.5-9B. It outperforms significantly larger models on Terminal-Bench 2.0 while maintaining high efficiency for local execution on mid-tier consumer hardware.
DISCOVERED
18d ago
2026-03-25
PUBLISHED
18d ago
2026-03-24
RELEVANCE
AUTHOR
Western-Cod-3486