Use on OpenRouter · Release blog post

Laguna XS 2.1-FP8

Laguna XS 2.1-FP8 is a 33B total parameter Mixture-of-Experts model with 3B 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 30 out of 40 layers for fast inference and low KV cache requirements.

[!NOTE] This is the FP8 variant with an FP8-quantized KV cache. The BF16, NVFP4 and INT4 variants are also available on Hugging Face.

Highlights

• Mixed SWA and global attention layout: Laguna XS 2.1 uses sigmoid gating with per-layer rotary scales, enabling mixed SWA (Sliding Window Attention) and global attention layers in a 3:1 ratio (across 40 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 33B total parameters and 3B activated, Laguna XS 2.1 is compact enough to run on a Mac with 36 GB of RAM. 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: 33B total with 3B activated per token
• Optimizer: Muon
• Layers: 40 layers (10 layers with global attention, 30 layers with sliding window attention)
• Experts: 256 experts with 1 shared expert
• Sliding Window: 512 tokens
• Modality: text-to-text
• Context window: 262,144 tokens
• Reasoning support: interleaved thinking with preserved thinking

Benchmark results

We used the highest publicly-referenced scores for all comparison models across each benchmark. In all cases these were official scores published in release blog posts or equivalent, with the exception of gpt-oss-120b and Claude Haiku 4.5 where the highest published (verified) scores for SWE-Bench Pro and Terminal-Bench 2.0 are from their respective official leaderboards.

Expand for benchmarking methodology

All benchmarking for Laguna XS 2.1 was completed using Laude Institute’s Harbor Framework with our agent harness, with a maximum of 500 steps and sandboxed execution. The same sampling parameters were used for all Laguna XS 2.1 benchmarking: temperature=1.0, top_k=20 and top_p=1, with thinking mode enabled and a context length of 256K tokens. All tasks were run in their own sandbox using 8 GB RAM/2 CPUs, with the exception of Terminal-Bench 2.0, which used 48 GB RAM/32 CPUs.

Some base task images and verifiers were patched to fix infrastructure reliability issues inherent in task setup, such as rate limits on third-party dependencies in external registries used by the verifier. All four agentic benchmarks were run with patched images. We also ran a reward-hack judge post-hoc on Laguna XS 2.1 evaluation runs and did not find significant reward hacking after joint judge review and manual review.

• SWE-bench Verified: mean pass@1 averaged over 4 attempts per task
• SWE-bench Multilingual: mean pass@1 averaged over 4 attempts per task
• SWE-Bench Pro: mean pass@1 averaged over 2 attempts per task
• Terminal-Bench 2.0: mean pass@1 averaged over 5 attempts per task; 48 GB RAM/32 CPUs

Quantization quality

Laguna XS 2.1-FP8 was evaluated against the unquantized BF16 checkpoint on the same four agentic benchmarks. Scores are mean pass@1; the ± figure is the run-to-run variation and RH is the share of runs flagged by our post-hoc reward-hack judge (see the benchmarking methodology note above).

Usage

Laguna XS 2.1-FP8 has launch-day support in vLLM, SGLang and Transformers, and TRT-LLM thanks to the support of the team at NVIDIA.

[!NOTE] For complete usage instructions, see the main Laguna XS 2.1 model card.

Local deployment

Laguna XS 2.1-FP8 is supported in vLLM, SGLang and Transformers, and TRT-LLM thanks to the support of the team at NVIDIA. Use Laguna XS 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

The full vLLM recipe is on the main Laguna XS 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-XS-2.1-FP8 substituted for the model ID. Set VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER=0 when serving with vLLM.

[!IMPORTANT] Tool calling requires a vLLM build with vllm-project/vllm#47311; on older builds pass --tool-call-parser glm47 (the GLM 4.7 parser) instead of poolside_v1.

[!NOTE] The FP8-quantized KV cache requires vLLM >= 0.22.0. Earlier versions produce scrambled output on non-Hopper GPUs because of a per-layer attention-head count bug, fixed in vllm#42650. On older vLLM, disable the FP8 KV cache by adding --kv-cache-dtype-skip-layers $(seq 0 39).

SGLang

Laguna XS 2.1 is supported in SGLang via sgl-project/sglang#24204. Quantization is detected automatically from quantization_config, so no extra flags are required. See the SGLang cookbook entry and the main Laguna XS 2.1 model card for a serving recipe.

Transformers

The full Transformers recipe is on the main Laguna XS 2.1 model card. Substitute poolside/Laguna-XS-2.1-FP8 for the model ID; quantization is detected automatically from quantization_config.

TRT-LLM

Laguna XS 2.1 support ships in TensorRT-LLM >=1.3.0rc16; see the install recipe on the main Laguna XS 2.1 model card. Substitute poolside/Laguna-XS-2.1-FP8 for the model ID; quantization is detected automatically from quantization_config, no extra flags required.

from tensorrt_llm import LLM

llm = LLM(model="poolside/Laguna-XS-2.1-FP8", trust_remote_code=True)

Ollama

Available on the Ollama library.

[!NOTE] macOS (Metal) users: Chat (ollama run / /api/chat) works as expected on Linux/CUDA. On macOS/Metal it may currently return empty output; the root cause is not yet fully understood and we're investigating it with the Ollama team. On a Mac, use a Linux/CUDA host, or the /api/generate endpoint with "raw": true.

Controlling reasoning

Laguna XS 2.1-FP8 uses the same reasoning controls (interleaved thinking, preserved reasoning, and the enable_thinking flag) as the base model. See the Controlling reasoning section of the main Laguna XS 2.1 model card.

License

This model is licensed under the OpenMDW-1.1 License.

Intended and Responsible Use

Laguna XS 2.1-FP8 is designed for software engineering and agentic coding use cases, and you are responsible for confirming that it is appropriate for your intended application. Laguna XS 2.1-FP8 is subject to the OpenMDW-1.1 License, and should be used consistently with Poolside's Acceptable Use Policy. We advise against circumventing Laguna XS 2.1-FP8 safety guardrails without implementing substantially equivalent mitigations appropriate for your use case.

Please report security vulnerabilities or safety concerns to security@poolside.ai.