Laguna XS 2.1 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. This model is an upgraded version of our Laguna XS.2 model with a +5.4% jump on SWE-bench Multilingual as well as stronger performance on terminal-style tasks.
Note. For more details on how we train, including on data automixing and async off-policy agent RL, check out our recent technical report.
Note. Laguna XS 2.1 is released under OpenMDW-1.1, a fully permissive license. Use it, modify it, and build commercial products on it. No permission required.
If you want more than the weights: production support, latency and cost optimization, or output indemnification, talk to us.
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. High-quality FP8, NVFP4 and INT4 quantized variants available (see the collection)
OpenMDW-1.1 license: Use and modify the model and associated materials freely for commercial and non-commercial purposes (learn more about OpenMDW)
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
Laguna XS 2.1 has launch-day support in vLLM, SGLang, Transformers and Llama.cpp, and TRT-LLM thanks to the support of the team at NVIDIA.
The fastest way to get started is using OpenRouter.
Note. We are providing free inference for a limited time for Laguna XS 2.1, as well as our larger 225B model, Laguna M.1. Visit our provider page on OpenRouter to get started.
Laguna XS 2.1 is supported in vLLM, SGLang, Transformers and Llama.cpp, and TRT-LLM thanks to the support of the team at NVIDIA. Use Laguna-XS 2.1 with Ollama (with MLX support) and the mlx-lm framework for the best experience on your local machine.
Serve Laguna XS 2.1 locally with vLLM and query it from any OpenAI-compatible client (see Controlling reasoning for tool calls, streaming, and reasoning extraction):
Note. Laguna XS 2.1 support is available in vLLM 0.21.0 and later (vllm-project/vllm#41129).
Important. Use a vLLM build that includes vllm-project/vllm#47311. Without that fix, the poolside_v1 tool parser silently drops Laguna XS 2.1 tool calls that have no newline after the function name, and the raw tool-call markup leaks into the response. On an older build, pass --tool-call-parser glm47 (the GLM 4.7 parser) instead.
Note.Optional: speculative decoding with DFlash. For lower latency, pair Laguna XS 2.1 with the DFlash speculator, a 5-layer Llama-style draft model that proposes up to 7 tokens per step at ~70% per-position acceptance on coding tasks. vLLM support is in progress in vllm-project/vllm#46853; once it lands, add --speculative-config '{"model":"poolside/Laguna-XS-2.1-DFlash","num_speculative_tokens":7,"method":"dflash"}' to the serve command above.
Note.Optional: speculative decoding with DFlash. The DFlash speculator can be paired with Laguna XS 2.1 for lower latency. SGLang support was added in sgl-project/sglang#29446. Add --speculative-algorithm DFLASH \ --speculative-draft-model-path poolside/Laguna-XS-2.1-DFlash-FP8 to the serve command above.
Laguna XS 2.1 support is merged into TensorRT-LLM (NVIDIA/TensorRT-LLM#13559) and ships in the v1.3.0rc16 pre-release wheels and later.
Note. A stable v1.3.0 has not been released yet, so install a pre-release wheel from NVIDIA's package index (the latest stable, v1.2.x, does not include Laguna XS 2.1 support). To build from source instead, see the TensorRT-LLM build documentation. Laguna XS 2.1 support is on main.
Install the CUDA-13 torch build first, then TensorRT-LLM. The default PyPI torch is a CUDA-12 build whose cuda-bindings pin conflicts with TRT-LLM's cuda-python 13.x, so a bare pip install tensorrt-llm fails to resolve; installing the cu130 torch up front avoids it.
# 1. CUDA-13 torch build (pins cuda-bindings 13.x, matching TRT-LLM's cuda-python)
pip install 'torch==2.10.0' torchvision --index-url https://download.pytorch.org/whl/cu130
# 2. TRT-LLM from NVIDIA's index (torch already satisfied, so it is not replaced)
pip install --pre 'tensorrt-llm>=1.3.0rc16' \
--extra-index-url https://pypi.nvidia.com \
--extra-index-url https://download.pytorch.org/whl/cu130
This resolves to tensorrt-llm 1.3.0rc20 with torch 2.10.0+cu130, cuda-python 13.0.3, and transformers 5.5.4.
Load the checkpoint directly with trust_remote_code=True. No transformers compatibility overlay is required: v1.3.0rc16+ pins transformers 5.5.4, which provides the symbols Laguna XS 2.1's config needs (earlier TRT-LLM releases pinned transformers 4.57, which did not).
from tensorrt_llm import LLM, SamplingParams
llm = LLM(
model="poolside/Laguna-XS-2.1",
trust_remote_code=True,
tensor_parallel_size=1,
)
sampling = SamplingParams(max_tokens=1024, temperature=1.0, top_k=20)
out = llm.generate(["Write a Python retry wrapper with exponential backoff."], sampling)
print(out[0].outputs[0].text)
The Laguna XS 2.1 tool-call and reasoning parsers are built into TRT-LLM >=1.3.0rc16 (shipped with #13559), so no extra install is needed. Note that the flag names differ from vLLM's (--tool_parser, and the reasoning parser is laguna, not poolside_v1).
The same recipe works for the FP8 and NVFP4 variants: quantization is detected automatically from quantization_config, no extra flags required.
Note.Optional: speculative decoding with DFlash. The DFlash speculator can be paired with Laguna XS 2.1 for lower latency. TRT-LLM support is in progress in NVIDIA/TensorRT-LLM#15666.
ollama run laguna-xs-2.1 # default — Q4_K_M (imatrix) ollama run laguna-xs-2.1:q8_0 # higher precision ollama run laguna-xs-2.1:bf16 # full precision
Reasoning and tool-calling work out of the box via the built-in laguna template.
Note.macOS (Metal) users:ollama run may currently return empty output on Apple Silicon (a Metal f16 overflow in the MoE down-projection; fix pending in ggml-org/llama.cpp#25389). Until it ships, use a Linux/CUDA host or the /api/generate endpoint with "raw": true.
Laguna XS 2.1 is also available in Atomic Chat, a desktop app for running local models with a simple chat UI. To try it, download Atomic Chat, open the app, and choose Laguna XS 2.1 from the recommended models.
Laguna XS 2.1 has native reasoning support and is designed to work best with preserved thinking, where reasoning content from prior assistant messages is preserved in the message history. This model will generally reason before calling tools and between tool calls.
Reasoning may not be generated in follow-up steps if prior thinking blocks are dropped (i.e., thinking is not preserved) when messages are reconstructed over multiple steps.
Expand for example
import json
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="...",
)
model = "poolside/laguna-xs-2.1"
tools = [{"type": "function", "function": {
"name": "shell",
"description": "Execute a bash command and return the output.",
"parameters": {"type": "object", "properties": {"cmd": {"type": "string"}}, "required": ["cmd"]},
}}]
messages = [
{"role": "system", "content": "You are a coding agent with access to a shell tool."},
{"role": "user", "content": "Run uname -a"},
]
# Thinking is enabled by default when the server sets --default-chat-template-kwargs {"enable_thinking": True}
# When using OpenRouter's Chat API (https://openrouter.ai/api/v1), this flag is set by default
response = client.chat.completions.create(
model=model,
messages=messages,
tools=tools,
stream=True,
)
reasoning, content, tool_calls = "", "", []
for chunk in response:
delta = chunk.choices[0].delta
if hasattr(delta, "reasoning_content") and delta.reasoning_content:
reasoning += delta.reasoning_content
if hasattr(delta, "content") and delta.content:
content += delta.content
if hasattr(delta, "tool_calls") and delta.tool_calls:
for tc in delta.tool_calls:
if tc.index >= len(tool_calls):
tool_calls.append({"id": tc.id, "function": {"name": "", "arguments": ""}})
if tc.function.name:
tool_calls[tc.index]["function"]["name"] = tc.function.name
if tc.function.arguments:
tool_calls[tc.index]["function"]["arguments"] += tc.function.arguments
print(f"Reasoning: {reasoning}\nContent: {content}\nTool calls: {tool_calls}\n")
# Return reasoning in the next request for best performance
messages.append({
"role": "assistant",
"content": content,
"reasoning_content": reasoning,
"tool_calls": [{"id": tc["id"], "type": "function", "function": tc["function"]} for tc in tool_calls]
})
messages.append({
"role": "tool",
"tool_call_id": tool_calls[0]["id"],
"content": json.dumps({"stdout": "Darwin arm64", "exit_code": "0"})
})
response = client.chat.completions.create(
model=model,
messages=messages,
tools=tools,
stream=True,
)
reasoning, content = "", ""
for chunk in response:
delta = chunk.choices[0].delta
if hasattr(delta, "reasoning_content") and delta.reasoning_content:
reasoning += delta.reasoning_content
if hasattr(delta, "content") and delta.content:
content += delta.content
print(f"Reasoning: {reasoning}\nContent: {content}")
You can disable thinking by setting enable_thinking to False in a request or by not providing --default-chat-template-kwargs {"enable_thinking": True} or equivalent when starting the server.
Expand for example
from openai import OpenAI
client = OpenAI()
completion = client.chat.completions.create(
model="poolside/laguna-xs-2.1",
messages=[
{"role": "user", "content": "Write a retry wrapper with exponential backoff."}
],
extra_body={
"chat_template_kwargs": { "enable_thinking": False },
},
stream=True
)
for chunk in completion:
print(chunk.choices[0].delta)
For agentic coding use cases, we recommend enabling thinking and preserving reasoning in message history as outlined in the Controlling reasoning section.
Laguna XS 2.1 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 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 safety guardrails without implementing substantially equivalent mitigations appropriate for your use case.
Please report security vulnerabilities or safety concerns to security@poolside.ai.
Files
29 files · 66.9 GBWeights 66.9 GB (safetensors). Sizes and sha256 from the signed manifest.
import os
os.environ["HF_ENDPOINT"] = "https://freeweights.allternit.com"
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("poolside/Laguna-XS-2.1", revision="c5f36269bbdbd3f27fddc9a9f9dbae0cf2cf57db")
model = AutoModelForCausalLM.from_pretrained("poolside/Laguna-XS-2.1", revision="c5f36269bbdbd3f27fddc9a9f9dbae0cf2cf57db", device_map="auto")
