SourcesPoolside
publisher · 5 sources
Poolside
5 sources read from this publisher, newest first, with the methods each one describes.
poolside/Laguna-S-2.1 · Hugging Face
first party release · model card · 2026-08-03 · 10 methods
Model release notes - Poolside
first party release · vendor docs · 2026-07-22 · 13 methods
Mixture of expertsThinking modeFP8 KV-cache quantizationSparse expert activationInterleaved thinkingAgentic edit-and-validate loop (read, edit, run checks)Checking for agent-generated test scripts before committingCheckpointing for rollbackIncremental task decomposition for agent promptsInstruction-following fine-tuningPlanning with the model for self-contained tasksRLCEFTask-specification prompting
Introducing Laguna S 2.1
first party release · official blog · 2026-04-28 · 23 methods
Mixture of expertsMax thinking mode with self-set compute budgetFP8-precision reinforcement learningSFT bootstrapping with synthetic dataTwo-stage post-training pipeline (SFT then RL)agentic repository installation taskautomated agentic research loop for software optimizationavg@kbenchmark-instrumented target code for agent optimizationEvaluation trajectory releaseHuman-calibrated LLM-as-a-judgeLong-horizon rollout budgetsModel FactoryMulti-harness rolloutsone-change-at-a-time benchmarking with measurable-wins retentionPrompt addendum against direct online solutionsReasoning effort configuration availabilityRL sandboxing with selective network blocking and artifact cachingsandboxed agentic benchmarking on a Harbor fork with step limitSeed-based terminal-task generationtraining objective for persistence and verification behaviorsTraining-task generation from real commit historyXML-tagged tool-call format
Two foundation models built for agentic coding.
first party release · vendor docs · 2025-10-28 · 2 methods
laguna m1 xs2 technical report
technical report · 116 methods
Mixture of expertsHybrid Sliding-Window and Global AttentionLong-context extension via continued pretrainingGrouped-query attentionMuonFP8 KV-cache quantizationShared expertGated DeltaNet and full attentionRotary Position EmbeddingBF16 mixed-precision trainingQuantization-aware distillationWarmup-Stable-DecayAutomatic training recovery after failuresKL-regularized data-mixture optimizationPersistent reasoning historyToken-in-Token-out (TITO)AutoMixerExponential moving average (EMA)GPU-to-inference weight sync over GPUDirect RDMAPatching benchmark images to remove git-history leaksSequence packingAgentic experiment workflowsAgentic RL task mix with a shared tool-use APIAtlas inference libraryAutomatic in-flight checkpoint evaluation schedulingAutomating repetitive research workAuxiliary-loss load balancingAWQ INT4 weight-only quantizationBehavioral-constraint augmentation with synthetic system messagesBinary task verifierBlind pairwise bucket-boundary calibrationBlocking in-flight rollout steps during weight updatesCapping trajectory stalenessCISPO with length-weighted leave-one-out group-relative advantagesClosed-loop multi-turn synthetic-data rolloutConfiguration-flag promotion of validated components to productionConservative model-based noise filteringCross-replica model-weight hash consistency checksCurrent-turn-only reasoning-mode detectionDagster