Laguna-S-2.1

Laguna S 2.1-NVFP4 is a 117.6B total parameter Mixture-of-Experts model with 8.5B activated parameters per token designed for agentic coding and long-horizon work on a local machine. It uses Sliding Window Attention with per-head gating in 36 out of 48 layers for fast inference and low KV cache requirements.

Updated release (August 2026). This is a new checkpoint that supersedes the earlier version of this repository. The weights have changed, not only the config, so if you downloaded a previous copy please re-download to pick up the current checkpoint.

Highlights

• Mixed SWA and global attention layout: Laguna S 2.1 uses softplus gating with per-layer rotary scales, enabling mixed SWA (Sliding Window Attention) and global attention layers in a 3:1 ratio (across 48 total layers)
• KV cache in FP8: KV cache quantized to FP8, reducing memory per token
• Native reasoning support: Interleaved thinking between tool calls with support for enabling and disabling thinking per-request
• Local-ready: At 117.6B total parameters and 8.5B activated, the NVFP4 weights are roughly 71 GB. Available on Ollama and llama.cpp (BF16 and Q4_K_M only)
• OpenMDW-1.1 license: Use and modify the model and associated materials freely for commercial and non-commercial purposes (learn more about OpenMDW)

Model overview

• Training: pre-training, post-training and reinforcement learning stages
• Number of parameters: 117.6B total with 8.5B activated per token
• Optimizer: Muon
• Layers: 48 layers (12 layers with global attention, 36 layers with sliding window attention)
• Experts: 256 experts with 1 shared expert
• Sliding Window: 512 tokens
• Modality: text-to-text
• Context window: 1,048,576 tokens
• Reasoning support: interleaved thinking with preserved thinking

Benchmark results

Benchmarks as of 21 July 2026. Laguna S 2.1 in bold; a dash (-) marks a benchmark a model was not evaluated on. Scores marked * are as reported by third parties: Terminal-Bench 2.1 and DeepSWE via Artificial Analysis, SWE Atlas via Scale AI's official leaderboard, and Toolathlon Verified via its official leaderboard. Full evaluation trajectories: trajectories.poolside.ai.

Usage

Context length

This checkpoint ships configured for a 1,048,576-token (1M) context window. The weights are native 1M checkpoints: training included a long-context extension stage up to 1,048,576 tokens, and quantization was calibrated at the 1M configuration.

If you prefer to cap the context at 262,144 tokens (256K), edit config.json:

"rope_parameters": {
  "full_attention": {
    "factor": 32.0,
    "attention_factor": 1.3465735902799727
  }
},
"max_position_embeddings": 262144

At long context you may experience some quality degradation.

Sampling

The sampling defaults in the checkpoint's generation_config.json are authoritative (top_k 20 is eval-certified truncation). Serve with those defaults rather than setting a separate temperature or top_p.

Local deployment

Laguna S 2.1-NVFP4 is supported in vLLM, Transformers, and TRT-LLM thanks to the support of the team at NVIDIA. NVFP4 does not currently run correctly on SGLang (see the SGLang note below). Use Laguna-S 2.1 with Ollama (with MLX support) or Llama.cpp (BF16 and Q4_K_M only) for the best results on your local machine.

vLLM

Serving requires vLLM 0.25.0 or later. See the main Laguna S 2.1 model card for the full recipe.

The full vLLM recipe is on the main Laguna S 2.1 model card and on the vLLM recipes page. Quantization is detected automatically from quantization_config in this checkpoint, so the same command works with poolside/Laguna-S-2.1-NVFP4 substituted for the model ID. No extra flags required.

[!NOTE] Optional: speculative decoding with DFlash. Pair with the quantization-matched draft model poolside/Laguna-S-2.1-DFlash-NVFP4 by adding --speculative-config '{"model":"poolside/Laguna-S-2.1-DFlash-NVFP4","num_speculative_tokens":7,"method":"dflash"}' to the serve command.

DGX Spark & Mac Studio

Ollama is the quickest way to run Laguna S 2.1 on a single 128 GB box, whether that is an NVIDIA DGX Spark (GB10) or an Apple Silicon Mac Studio:

ollama run laguna-s-2.1

It pulls the Q4_K_M GGUF (about 75 GB) and runs on the GB10 GPU on the Spark or on Metal on the Mac, with no build step. We measured about 12.6 tokens/s on the Spark and 17.6 on Apple Silicon.

• You need about 128 GB of unified memory; the Q4_K_M weights are around 75 GB on their own.
• The first load reads 75 GB off disk and can run past Ollama's 5-minute default timeout. If it does, set OLLAMA_LOAD_TIMEOUT=20m.
• Reasoning is on by default. Pass "think": false in the request to turn it off.
• The higher-fidelity tags need a bigger box: `laguna-s-2.1:q8_0