TAAFT
Free mode
100% free
Freemium
Free Trial
Deals
Companies
poolside
Models
Laguna M.1

Laguna M.1

By poolside
Laguna M.1 is a 70-layer MoE transformer with 225B total parameters and 23B activated per token. The first 3 layers are dense SwiGLU; the remaining 67 are sparse MoE with 256 experts plus 1 shared expert, top-k=16 routing, and auxiliary-loss-free load balancing. It uses global attention across all layers with 64 Q-heads, 8 KV-heads, head dimension 128, and softplus attention output gating. Positional encoding uses RoPE with YaRN, supporting a 262,144-token context window. Trained on 30T tokens using the Muon optimizer through pre-training, post-training, and async off-policy agent RL stages on 6,144 NVIDIA Hopper GPUs. Features native reasoning with interleaved thinking between tool calls, configurable per-request. Benchmark results: 74.6% SWE-bench Verified, 63.1% SWE-bench Multilingual, 49.2% SWE-bench Pro, 45.8% Terminal-Bench 2.0. Supported in vLLM, SGLang, Transformers, and TRT-LLM.
New Coding Gen 2
Released: May 26, 2026

Overview

Laguna M.1 is a 225B total parameter (23B activated) Mixture-of-Experts model built for agentic coding and long-horizon software engineering. It features a 262,144-token context window, native interleaved reasoning with per-request thinking control, and tool calling support. Achieves 74.6% on SWE-bench Verified and 49.2% on SWE-bench Pro. Apache 2.0 licensed.

‍💻Code generation📝Text generation🤖Ai agents

About poolside

poolside is a foundation-model company building AI for software engineering. It develops proprietary frontier models (Malibu, Point, and the open-weight Laguna XS.2) plus a full agentic platform — IDE/TUI integrations, multi-agent systems, and connectors — deployed on-prem, in customer VPCs, or in air-gapped environments. Targets large enterprises and the U.S. defense industrial base. Mission stated as building AGI via software engineering as the strategic beachhead.

Industry: Artificial Intelligence
Company Size: 322
Location: San Francisco, California, US
Website: poolside.ai
View Company Profile

Related Models

  • Gen 2
    GPT 5.3 Codex
    GPT-5.3-Codex is OpenAI’s most capable agentic coding model, combining GPT-5.2-Codex’s frontier coding with GPT-5.2’s reasoning in one model that is ~25% faster, tops SWE-Bench Pro and Terminal-Bench, and runs long, tool-using workflows on a computer.
    💻Coding🔍Code reviews🔎Problem solving📝Writing
    Coding
    Released 4mo ago
  • Gen 2
    CSM
    Conversational speech generation model that generates audio codes from text and audio inputs for dialogue style speech output.
    🔊Text to speech
    Coding
    Released 4mo ago
  • Gen 2
    Voyage code 3
    Text embedding model optimized for code retrieval. Supports Matryoshka embeddings at 256, 512, 1024, and 2048 dimensions with quantized formats (float, int8, uint8, binary, ubinary) for up to 32x storage reduction. Features a 32K-token context window. Outperforms OpenAI text-embedding-3-large by 13.80% on 32 code retrieval datasets covering text-to-code, code-to-code, and docstring-to-code tasks.
    🔎Semantic search🔍Code search📦Code retrieval
    Coding
    Released 1y ago
Last updated: June 19, 2026
0 AIs selected
Clear selection
#
Name
Task