Research, case studies, and engineering deep-dives from the Neo team.
NEO ran an autonomous research loop on the pi coding agent and raised QuixBugs from 49/60 to 57/60 on development, then 19/20 on a sealed holdout. Same model. Five prompt guidelines. One file changed.
On 38 comparable answers, Muse Glimmer 30B wins deterministic checks while a blind LLM judge prefers Qwen3.8-27B. Built and run end to end with Neo — the winner depends on how you grade.
Audio8 now clones on CPU: identity near-tie with XTTS (SIM 0.5454 vs 0.5569). Among cloners Audio8 is the practical pick (UTMOS 3.3657, RTF 1.4214); Kokoro still wins fixed-voice quality and speed. VCTK sweep built with Neo.
Grok 4.6, Kimi K3, and Opus 5 each tried to speed up Kokoro-82M TTS on CPU. Study scores: Opus 83, Kimi 60, Grok 51. Grok's best verified gain was about 6–8% in one setup, but audio checks failed, so nothing shipped.
OmniVoice sounded good but felt slow. Without training a new model, we made it about 6× faster than the default path, and still about 3× faster than the official fast setting, then kept the quality checks honest.
Five ASR models on 296 real clips across 19 conditions. Clean-speech WER ties at 4.4%, then accent and noise separate the field, with committed transcripts, bootstrap CIs, and a CI regression gate.
Using Neo BYOK, both models optimized Kokoro-82M on CPU. Opus built the stronger study, but its claimed 18.66% speedup became 9.81% slower on replay.
Three AutoML artifacts, all green suites (45/45, 83/83, 37/37). Execution found silent trust failures. Scores: Kimi 68, Fable 61, GLM 54.
Both shipped green suites; execution inverted the ranking. Kimi K3 68/100 vs GLM 5.2 54/100 on production ML frameworks. Built with NEO BYOK.
Neo profiled NVIDIA Parakeet TDT 0.6B v3 on CPU, ran keep/discard ladders, and froze a static-QDQ production pack that cut primary RTF by ~2.07× on EPYC (~1.42× on Apple Silicon). Runtime-only knobs never cleared 5%.
Kyutai's Pocket TTS joins the CPU TTS benchmark: 6 configs, 180 timed runs, and 36 WAV samples across RTF, latency, throughput, and UTMOS MOS, plus zero-shot voice cloning from 5 seconds of audio. Built end-to-end with Neo.
Neo evaluated Qwythos-9B at Q4_K_M and Q8_0 on GSM8K, IFEval, and HumanEval from a single prompt. GSM8K hit 84%, IFEval 66%, HumanEval 0% — and Q4 is nearly as good as Q8 for math.
