Axolotl v0.16 vs Structured Output Benchmark

Axolotl is the go-to open-source fine-tuning framework for the local LLM community, and v0.16 is its most significant performance release to date. The headline numbers are striking: 15x faster training for Mixture-of-Experts (MoE) models with LoRA adapters, 40x reduction in memory usage for the same configurations, and 58% faster GRPO async training — the algorithm behind many of the recent reasoning model breakthroughs. Day-0 support for Google Gemma 4 shipped simultaneously with the model release. The MoE+LoRA improvements are especially timely. As sparse mixture-of-experts models like Gemma 4, Mistral, and Qwen3.6-Plus dominate the model landscape, fine-tuning them has been disproportionately expensive. Axolotl v0.16 makes it practical to fine-tune these architectures on a single consumer GPU — previously a multi-GPU or cloud-required task. The GRPO improvements also make reinforcement learning from human feedback (RLHF) workflows dramatically faster for small teams. For the indie fine-tuning community — researchers, small companies, and hobbyists building specialized models — this release removes a major cost barrier. Combined with the simultaneous Gemma 4 support, v0.16 positions Axolotl as the fastest path from a new model release to a fine-tuned, production-ready custom variant.

Interfaze's Structured Output Benchmark (SOB) exposes a gap that has been quietly breaking production AI pipelines: models can produce syntactically valid JSON while getting the actual values wrong. SOB measures value accuracy across 21 models using 5,000 text passages, 209 OCR documents, and 115 meeting transcripts — scoring each on seven metrics including value accuracy, faithfulness (grounding vs. hallucination), type safety, and perfect-response rate. The benchmark reveals some sobering findings. Even top models like GPT-5.4 and Claude Sonnet 4.6 achieve ~83% on text but drop to 67% on images and only 23.7% on audio. No single model dominates all modalities — GPT-5.4, GLM-4.7, Qwen3.5-35B, and Gemini 2.5 Flash cluster within one point of each other on text. Perfect response rates (all seven metrics correct) rarely exceed 50% for even the best performers. For developers building data extraction pipelines, agents that read invoices, or any system where "correct JSON" means more than syntactically valid JSON, this is required reading. The dataset is on Hugging Face, the paper is on arXiv, and the playground lets you test your own model's structured output capability directly.