Rendered
Benchmarks and engineering notes from Sprag.
Technical writing on multimodal inference, serving systems, and the work behind production AI infrastructure.
Speech-to-Speech vs the Traditional Voice Pipeline: The cost of Modularity
Transcription throws away tone, emphasis, and emotion before they reach the next stage of a voice pipeline. In our agentic benchmarks, Symphony finishes a turn 500 ms faster than our own Rhythm + Chorus pipeline.
Alex Smith
CEO
Recent
More from Rendered
A Transcript Is a Choice
Symphony nearly matches a purpose-built verbatim transcription model on spoken disfluencies, and can turn the same recording into the representation each part of a voice agent needs.
Qwen3.8-27B is free on Sprag through the end of August
Qwen3.8-27B is open weights under Apache 2.0 and built to be run locally. Running it locally still takes a 24GB card, so we're serving it for free through the end of August.
How ISSEN delivers AI language tutoring at scale with Sprag
ISSEN (YC F24) replaced a deprecated Gemini Flash 2.0 transcriber with Qwen3-Omni served on Sprag, reaching sub-300ms multilingual STT latency and 90% accuracy on hard code-switched evals.
Multimodal API compatibility from the adapter layer
Implementation notes from building compatibility across OpenAI, OpenRouter, Hugging Face, Cartesia, and ElevenLabs. Request shape, response shape, and usage units should be treated as three separate problems.
Voice Agent Economics in 2026
Native speech-to-speech now runs 3–12× cheaper than cascaded STT → LLM → TTS stacks at comparable latency and task success. Here's when to pick native, when cascades still win, and what changed.
Speech-to-Speech vs Pipelined Voice Agents
We benchmarked Qwen3-Omni against traditional ASR → LLM → TTS pipelines on real voice-agent workloads. For the first time, the open native path actually looks preferable.
Introducing Sprag: Unified Multimodal AI Inference
Today we're launching Sprag — a single API for text, image, audio, and video inference. One endpoint, every modality, built for speed.