Baseten touts Mirendil base model's record-low perplexity; giffmana questions the data
giffmana · x · 2026-10-24
Baseten announced it ran Mirendil's new base model on its private codebases — code the model had never seen — reporting the lowest bits per byte (a tokenizer-independent perplexity measure) of any model tested, and signed on as inference partner for the launch. giffmana pressed back asking whether Baseten's codebases truly contain no vLLM, SGLang, or TensorRT-LLM code, challenging the "never seen it" framing.
Related event: Baseten Benchmarks Show Mirendil-Base-Large Beats Rivals on Perplexity(3 posts)→
More from Models
- Nvidia and UW Institute for Protein Design open-source RFD4-Proteina protein design model — AllThingsApx · 2026-10-24
- Local quant benchmark: new releases trail months-old DeepSeek/Qwen models — EmilPi · 2026-10-24
- Meta's Muse Spark Open hits ~150M tokens/s per MW on NVIDIA GB300 in early benchmarks — ArtificialAnlys · 2026-10-24
- Kimi K3.1 falls short of benchmarks in hands-on testing, says KOL — DeryaTR_ · 2026-10-24
- Kimi K3.1 Pro tops agentic coding benchmarks but may lag on AA-3D — ArtificialAnlys · 2026-10-24
- OpenAI says it will stop training models on the live internet, citing risks — TechCrunch AI · 2026-10-24