Evidence
- Aug 20Claimed to beat same-size Qwen 3.5 and approach or edge Qwen's 35B MoE on several agentic and reasoning benchmarks
- Aug 20Served with vLLM on an 80GB A100 used about 72GB VRAM, mostly KV cache; lowering context length reduces it
- Aug 20Failed an AWS CLI EC2 provisioning task in Hermes Agent: command errors, typos, 43k tokens in 7 minutes without finishing
- Aug 20Struggled on a single-file HTML rotisserie chicken coding task, producing garbage and hallucinating in Chinese; reviewer calls it unusable
Cost to run
Alternatives
Made from