Xiaohongshu open-sources Iris search-agent model, hitting 82.2 on BrowseComp
大模型之路 · wechat · 2026-09-23
Xiaohongshu's AllSpark team open-sourced Iris, a search-agent model with fully transparent weights, eval code, and context-management — the first privately deployable open base for DeepResearch-style agents.
- Specs: two MoE tiers — Iris-mini (35B total/3B active) and Iris-pro (397B/17B), both with 256K context. The model internalizes when to search, how to read, and when to stop, unlike RAG pipelines.
- Data: automatic multi-hop question synthesis over web hyperlink graphs, with entities rewritten to descriptive references to block keyword-matching shortcuts; dual filtering (unsolvable closed-book, solvable open-book).
- Training: SFT-RL climbing; reward judging and summarization kept inside the training cluster with no external APIs; long trajectories resumed at request level.
- Results: Iris-mini scores 82.2 on BrowseComp, nearing trillion-parameter models like Kimi-K2.6; skills spill over into tool calling (BFCL, τ-bench).
- Key finding: without context management BrowseComp drops to 64.7; clearing accumulated tool history adds 17.5 points. Iris-mini runs locally at 4-bit on 30GB VRAM.
More from coding & agent
- Packaging a local MCP server as a one-click Claude Desktop extension: lessons learned — sarghayk · 2026-10-24
- Cloning Your Boss: Distilling a Year of Slack Into an AI Persona File — every · 2026-10-24
- Replit Agent clones iOS Stocks app that runs natively-smooth on the web — amasad · 2026-10-24
- Building landing pages with MiniMax M3.1 Flash at $0.15/M tokens, prompts included — gaganghotra_ · 2026-10-24
- Kornia 0.9.0 ships a conventions page targeting LLM coding pitfalls — kornia_foss · 2026-10-24
- Alchemy Merges All-in Migration to EffectTS Native SQL Drivers, Dropping Raw pg/mysql2 Deps — samgoodwin89 · 2026-10-24