Specialist-model research desk

Saved workspace
Comparisons

Which option should we test first?

Frontier Development research deskSection 03 · Page 1 of 11Last updated 26 October 2026

Choose the option that addresses the diagnosed failure with the least irreversible work. Compare architectures separately from who delivers them: a managed platform, a consultancy and an internal team can implement similar methods with very different responsibilities.

The tables below are the desk’s decision framework. They are not a scored vendor league table, and “fits” describes a hypothesis to test. Use one workload, one acceptance specification and a dated baseline throughout.

TRACE THE EVIDENCE

Section source library

This page contains no direct external citations. These are the section’s references; they do not independently validate the desk’s proposed scoring or methods.

  1. A Small Model Asks Better: Fine-Tuning Query Clarification for Retrieval-Augmented Generation — He, Popa and Danilevsky, 19 October 2026 →arxiv.org
  2. Together AI — AI Native Cloud →www.together.ai
  3. Llama 3.3 Community License Agreement — Meta →github.com
  4. Pricing — Together AI, snapshot checked 26 October 2026 →www.together.ai
  5. Demystifying evals for AI agents — Anthropic, 9 January 2026 →www.anthropic.com
  6. SFT Trainer — Hugging Face TRL documentation →huggingface.co
  7. RouteLLM: Learning to Route LLMs with Preference Data — Ong et al., revision 23 February 2025 →arxiv.org
  8. Extend — Extraction Overview →docs.extend.ai
  9. QLoRA: Efficient Finetuning of Quantized LLMs — Dettmers et al., 23 May 2023 →arxiv.org
  10. LoRA Learns Less and Forgets Less — Biderman et al., revision 20 September 2024 →arxiv.org
  11. Quantization — Hugging Face PEFT documentation →huggingface.co
  12. Fireworks — specialised intelligence platform →fireworks.ai
  13. Winder.AI — Generative AI Consulting & Development Services →winder.ai
  14. Modal — platform for production AI →modal.com
  15. Baseten — inference platform →www.baseten.co