INT21 generates Rust inference runtime for Drama 3 in 1 hour, 63% faster than SGLang
bingxu_ · x · 2026-10-15
INT21 says it built a Rust inference runtime for Fish Audio's open-source Drama 3 speech model, generated by AI in 1 hour.
- vs SGLang-Omni Day 0 support: time-to-first-audio cut 38-60% (15.11ms vs 37.79ms on "Hello"), audio generation throughput up 59-63% (223.4 vs 136.9 frames/s)
- Benchmarked on 2× NVIDIA B200 per runtime (one for generation, one for codec), single speech stream, 1-frame first chunk, 20 warm trials per prompt
- Fish Audio also released the Drama 3 open-source speech model
More from coding & agent
- sgrep cuts coding agent LLM costs 21.3% on SWE-bench versus ripgrep, fully local in Rust — udmrzn · 2026-10-15
- 10 GitHub repos that turn Claude Code and Codex into elite UI designers — udmrzn · 2026-10-15
- Grep adds 1,000+ Composio integrations and authenticated browser access for agents — ycombinator · 2026-10-15
- Cloudflare's Clef decision models land on Merge Gateway, median latency 39ms — shensi · 2026-10-15
- Mine Bittensor Subnet 92 by pointing Claude Code at one link — airesearch12 · 2026-10-15
- Future agents will keep refactoring memory, building RL environments, and distilling to old checkpoints — manosaie · 2026-10-15