Baseten & Harvey Discuss Post-Training for Long-Context Open Models
baseten · x · 2026-07-23
Baseten reshared insights from the AI legal platform Harvey's engineering team. Harvey noted that for long-horizon tasks in the legal domain, even massive context windows of 256k to 1M tokens are often insufficient.
To address this, they discussed technical solutions like KV cache compaction. The teams also shared their data strategies for long-horizon tasks and best practices for effectively injecting domain knowledge into open-source models.
More from coding & agent
- A code sketch imagines an agent that upgrades itself and swaps models on demand — irvinebroque · 2026-07-23
- Claude Code plus Pika’s Voxel-it skill turns a meetup photo into a Minecraft scene — ytjessie_ · 2026-07-23
- A $20 Ethernet link is enough for 39.7GB multi-node GPU inference — Chuyito · 2026-07-23
- OpenAI GPT-5.6 Sol helped solve 6 open Erdős problems in 5 days — Charuru · 2026-07-23
- ModelScope launches an MCP server for text-to-image generation with Qwen-Image — modelcontextprotocol · 2026-07-23
- AILANG Parse brings deterministic DOCX, PPTX, XLSX and PDF parsing to MCP — modelcontextprotocol · 2026-07-23