PublishersPoolside

Poolside

5 documents read from this publisher.

poolside/Laguna-S-2.1 · Hugging Face

first party release · model card · 2026-08-03 · 9 techniques

Mixture of Experts Hybrid Attention Grouped-query attention DFlash Preserved thinking history mode Interleaved thinking between tool calls Per-layer-type rotary position scales Softplus attention gating Token-choice routing with softplus gating

Model release notes - Poolside

first party release · vendor docs · 2026-07-22 · 13 techniques

Mixture of Experts Configurable reasoning effort FP8 KV-cache quantization Sparse expert activation Interleaved thinking between tool calls Agentic edit-and-validate loop (read, edit, run checks)Checking for agent-generated test scripts before committing Checkpointing Incremental task decomposition for agent prompts Instruction-following fine-tuning Planning with the model for self-contained tasks Reinforcement Learning from Code Execution Feedback Task-specification prompting

Introducing Laguna S 2.1

first party release · official blog · 2026-04-28 · 23 techniques

Mixture of Experts avg@k FP8-precision reinforcement learning Reinforcement learning for low-pass-rate tasks SFT bootstrapping with synthetic data agentic repository installation task automated agentic research loop for software optimization benchmark-instrumented target code for agent optimization Full trajectory release Human-calibrated LLM-as-a-Judge Long-horizon rollout budgets Model Factory Multi-harness rollouts one-change-at-a-time benchmarking with measurable-wins retention Prompt Addendum Prohibiting Direct Online-Solution Use RL sandboxing with selective network blocking and artifact caching sandboxed agentic benchmarking on a Harbor fork with step limit Seed-based terminal task generation Thinking mode with automatic test-time compute budget training objective for persistence and verification behaviors Training-task generation from real commit history User-configurable thinking-effort control XML-tagged tool-call format

Two foundation models built for agentic coding.

first party release · vendor docs · 2025-10-28 · 2 techniques

Harbor Reinforcement learning in an agent harness

laguna m1 xs2 technical report

technical report · 115 techniques

Mixture of Experts Hybrid Attention FP8 KV-cache quantization Grouped-query attention Muon Shared Experts Rotary Position Embedding Warmup-Stable-Decay Continual pretraining for long-context extension Cross-turn persistent reasoning history Token-in-token-out (TITO)512-token Sliding Window Attention AutoMixer Removing Git-History Leaks from Benchmark Images Agentic RL Task Mix Agentic Workflows Atlas inference library Automatic in-flight checkpoint evaluation scheduling Automating Repetitive Infrastructure Work Auxiliary-loss load balancing AWQ INT4 weight quantization (W4A16)BF16 mixed-precision training Binary task verifier Binary terminal-verifier reward Blind pairwise bucket-boundary calibration Blocking in-flight rollout steps on weight updates Capping trajectory staleness CISPO with Length-Weighted Leave-One-Out Group-Relative Advantages Closed-loop multi-turn rollout Configuration-flag promotion of validated components to production Conservative model-based noise filtering Continuous contribution-score ranking Cross-replica model-weight hash consistency checks Current-turn-only reasoning-mode detection DAG-based asset lineage tracking Decompositional instruction-following judge Dense annotation of ambiguous low-quality data Dense gated attention with full RoPE and full gating Deterministic chain of checkers Diverse benchmark validation for quantization