Slay the Spire: the full fine tuning report
Nimble-9B, a 9B LLM, and Laya english, a 421M encoder, run on one machine and play the same seed with the same text and options. The full report shows what fine tuning changed.
INSIGHTS
Ideas behind the infrastructure.
Explore the engineering principles and research directions shaping our work.
MINDRAI / X
@Mindr4AIResearch in the open.
View original on X
Apollo 11 radio meets open ASR
An open speech recognition experiment with Apollo 11 landing radio and a 59 minute press conference, exploring transcription, speaker identification, and timestamps.
Local models play Slay the Spire
Fine tuned nimble-9b and laya english models run entirely locally and compete in the strategy card game Slay the Spire.
GPT Image 2.5 vs. Qwen-Image-2.1
A comparison of cloud and local image generation, examining face and outfit consistency across close up and full body views.
Ten images, one composition
Qwen-Image-2.1 runs locally with model-compose to combine separate images into one composition, with notes on DGX Spark, RTX 4090, and MacBook M2.
YuE2-3B: vocals and instruments together
An introduction to an open music model that generates stereo audio from a style prompt and lyrics, with score planning or direct audio generation.
YuE2 and Suno, side by side
A music generation comparison using the same lyrics and style, with YuE2 running on a local machine alongside Suno.
BUILD WITH ASTARIA
Start a conversation