Poolside releases Laguna S 2.1, a 118B MoE coding model
Poolside says Laguna S 2.1 is its most capable model yet: a 118B-parameter Mixture-of-Experts system with 8B parameters activated per token, up to 1M-token context, and both thinking and no-thinking modes. The company positions it as competitive with much larger models for coding and agent workloads.
Key takeaway
A mid-tier coding lab is shipping a sparse MoE with long context and dual inference modes, tightening competition below the frontier giants without needing dense 100B-plus activation costs.
Context
Poolside announced Laguna S 2.1 on X as an official release, citing 118B total parameters, 8B active per token, a context window of up to 1 million tokens, and optional thinking versus non-thinking generation. That combination targets software engineering and agent use cases where long context and controllable compute matter more than raw dense size.
An @llmgram quote flagged the marketing language around MoE versus dense activation, but the underlying event remains a first-party model launch with named specs. For an AI news channel, the material news is the shipping of Laguna S 2.1 itself, not the editorial quibble over how the architecture is described.
Numbers to know
- 118Btotal parameters claimed for Laguna S 2.1
- 8Bparameters activated per token
- 1Mmaximum context window in tokens