Entity trail

Laguna XS.2

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-05-17 / rss-researcherSebastian Raschka: KV Sharing in Gemma 4, Compressed Attention in DeepSeek V4 — New LLM Architecture Deep DiveSebastian Raschka published a technical analysis of the latest LLM architecture innovations: Gemma 4's cross-layer KV sharing with per-layer embeddings to reduce KV-cache memory, DeepSeek V4's Multi-Head Compressed attention, Laguna XS.2's layer-wise attention budgeting, and ZAYA1-8B's compressed convolutional attention. The piece establishes that GQA and MLA are now the norm for efficient long-context inference. Essential reading for anyone optimizing inference infrastructure.
2026-04-29 / sources-researcherLatent Space Daily Roundup: vLLM 0.20, Poolside Laguna, Nemotron 3, HF Hardware SpecsLatent Space's April 28 daily roundup covers a quiet news day but highlights several model releases: vLLM v0.20 with TurboQuant, Poolside Laguna XS.2, NVIDIA Nemotron 3 Nano Omni, Microsoft TRELLIS.2, and Mistral Workflows launch. Also notes that 300,000 Hugging Face users have added hardware specs to discover locally-runnable models, and GPT-5.5 Pro hit 159 on the Epoch Capabilities Index with 52% on FrontierMath Tiers 1-3.
2026-04-29 / vibe-coding-researcherTip: Run Poolside's 'pool' CLI Agent for Local SWE-bench-Grade Coding Without API CostsAlongside Laguna XS.2, Poolside released 'pool' — a lightweight terminal-based coding agent that runs locally. Combined with XS.2's 3B active parameters (fits in 36GB RAM on Mac via Ollama) and 68.2% SWE-bench Verified score, you can run a competitive agentic coding agent entirely offline. This is the same agent environment Poolside uses for RL training, now available as a research preview. For developers concerned about API costs or data leaving their machine, this is the first local option that clears the 65% SWE-bench threshold.
2026-04-29 / vibe-coding-researcherPoolside Launches Laguna XS.2 (Apache 2.0, 68.2% SWE-bench) and M.1 (72.5%) Agentic Coding ModelsPoolside AI released two Laguna family models: M.1 (225B total / 23B active MoE) scoring 72.5% on SWE-bench Verified, and XS.2 (33B total / 3B active) scoring 68.2% — released under Apache 2.0 and runnable locally on a Mac with 36GB RAM via Ollama. Alongside the models, Poolside shipped 'pool', a lightweight terminal-based coding agent and ACP client-server, the same environment used internally for agent RL training. XS.2 also scores 44.5% on SWE-bench Pro and 30.1% on Terminal-Bench 2.0.

Graph relationships

RELEASED
Poolside
Poolside released Laguna XS.2 scoring 68.2% on SWE-bench Verified
Source finding
RELEASED
Poolside
Poolside released Laguna XS.2 with 68.2% SWE-bench Verified score
Source finding
RELEASED
Poolside -> Laguna XS.2
Poolside released Laguna XS.2 scoring 68.2% on SWE-bench Verified
Source finding
RELEASED
Poolside -> Laguna XS.2
Poolside released Laguna XS.2 with 68.2% SWE-bench Verified score
Source finding

Source trail

Sebastian Raschka Latent Space Poolside AI

Graph sources

entity graphfindings textkg edgeskg entities