The author fine-tuned a 0.8B parameter model called SpeakoFlow Mini for dictation cleanup, which matched GPT-5.6 Luna on a specialized benchmark under fixed prompt conditions.
I built SpeakoFlow Mini, an Apache-2.0 fine-tune of Qwen3.5-0.8B for dictation cleanup. It is not a chat model or a general rewriter. It takes speech-to-text output, applies corrections the speaker actually made, and leaves everything else alone. That last part is harder than it sounds. General models often polish text that was already correct, but in dictation an unnecessary improvement is still a wrong edit. For example: Input: The deadline is Monday. Scratch that. The deadline is Wednesday. Output: The deadline is Wednesday. On my specialized English-only benchmark, SpeakoFlow Mini scored 70.7% and GPT-5.6 Luna scored 65.0%. Both used the same fixed short prompt, with reasoning disabled. The gap was +5.8 points, but the 95% interval was [-1.5, +12.9], so this is a statistical tie, not a win. That setup matters. It is the low-overhead configuration this model was trained for. With a longer prompt and a reasoning budget, Luna does better. It is also the better model for unusual cases outside the narrow patterns covered here. The claim is narrower. Under one fixed short prompt with no reasoning budget, a 0.8B model running locally matched a hosted frontier model on this task. The controlled comparison is against the untuned Qwen3.5-0.8B base. Fine-tuning moved the score from 47.3% to 70.7%, a gain of +23.4 points with a 95% interval of [+16.3, +30.3]. The second image shows that comparison. The Q8_0 build is 833 MB and runs fully offline. The Hugging Face card has the model files, run commands, benchmark method, quantization results, limitations and public examples: https://huggingface.co/SpeakoFlow/speakoflow-mini The evaluation set is not part of the public release. If you try it, I am most interested in cases where it changes something that should have been left alone. Disclosure: I built SpeakoFlow Mini and SpeakoFlow.
Built and released BetterGPT-150M, a compact 150M parameter causal language model that outperforms GPT-2 Small with low resource footprint. Includes live Hugging Face Space demo for text completion.
Independent study shows 227M-parameter hypernetwork adds zero gain over well-crafted few-shot prompts for tool-use in 3B Llama, achieving 79.7% of GPT-5 performance at 10× lower latency.
superwhisper/s1-mini is a 0.6B-parameter text normalizer fine-tuned from Qwen3-0.6B to clean speech-to-text transcripts by removing fillers, correcting errors, and applying punctuation and formatting, achieving 94.8% accuracy on English data.
A new toolkit enables fine-tuning the tiny Inflect-Nano/Micro TTS models on custom voice and language, supporting warm-start, resumption, and export to PyTorch/ONNX.
OpenAI demonstrates fine-tuning GPT-2 (774M parameters) using human preference feedback for text continuation and summarization tasks, requiring 5k labels for stylistic tasks and 60k for summarization, with models achieving 86-88% human preference rates though revealing labeler heuristic exploitation.