🤖IQuest-Q1: Open 320B MoE Model Built for Agentic CLI
TL;DR
IQuest-Q1 is an open sparse MoE with 320B total and 15B active parameters, tuned for agentic coding. The paper details a three-stage pipeline but reports no benchmark scores in the abstract.
IQuest-Q1 is an open sparse MoE with 320B total and 15B active parameters, tuned for agentic coding. The paper details a three-stage pipeline but reports no benchmark scores in the abstract.
Key Points
320B total parameters, 15B activated per token
Pre-training uses Stable SplitNorm, Muon and muP with a coding-heavy second stage
Post-training mixes SFT, domain RL and MOPD across several agent scaffolds
Model merging combines experts from different stages
No benchmark numbers in the abstract, so treat claims as unverified
Why It Matters
Another open coding-agent MoE at 15B active params is cheap to serve. Wait for independent evals before you switch.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,567 builders reading daily.