Task-specific NLP comeback: ScaleDown's CDM claims 200x speed, 467x cost savings
量子位 · wechat · 2026-10-18
Task-specific small language models are having a moment. ScaleDown launched CDM (constrained decoding), a trio of models for classification, entity extraction, and summarization that skip token-by-token generation and return structured results directly — claiming up to 200x speed and 467x cost savings vs frontier APIs, 1M-token text/image input, and pricing of $0.05 per million input tokens with free output. Earlier ScaleDown benchmarks showed 8-9% higher accuracy than Claude/GPT/Gemini at a fraction of the cost; the company has raised $24.5M led by Khosla Ventures.
Rival Jev (TypeSafe AI) shipped v1.14.0: multimodal image input (actually present since 1.13), a Box output type for object localization, multilingual improvements, and experimental long context. On DecisionBench, CDM scores 77.2% vs Jev's 72%. The trend: as AI apps scale, cost, latency, and output stability are reviving classic NLP tasks as dedicated model services.
More from Models
- ChatGPT web can actually listen to audio, but the same model can't via API or Codex — Possible_Statement84 · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- TLA+ expert corrects Grok: TLC only exhausts finite state spaces, not all behaviors — tianyin_xu · 2026-10-23
- Open LLM cost calculator shows Arabic tokenization and 6x price gaps across models — Moham-saed · 2026-10-23
- OpenAI blocks manual reset with 13% usage left: 'no reset needed right now' — AccomplishedSugar490 · 2026-10-23
- Meta's 816B Muse Spark Open scores 39, losing to 180B Qwen3.8-Flash-Next at 40 — BTA_Labs · 2026-10-23