The 3B parameter model that beats 137B on coding benchmarks and runs on a laptop

The number that catches the eye in Poolside's Laguna XS 2.1 announcement is 5.4. That's the gain on SWE-bench Multilingual, up to 63.1%. But the real story is not about any single benchmark delta. It's about what happens when a capable coding model fits into 3B active parameters and comes with the infrastructure to actually run it on a laptop. Efficiency matters more than raw size, as Nvidia's embedding models show.

XS 2.1 uses the same Mixture-of-Experts architecture as its predecessor XS.2: 33B total parameters, 3B activated per token. The improvement comes from better training, not a bigger model. The company says the improvements are concentrated on terminal-style tasks and multilingual coding, areas where the original XS.2 was weaker.

The benchmarks that matter for local use

The table Poolside published shows XS 2.1 at 63.1% on SWE-bench Multilingual, compared to XS.2 at 57.7%. On SWE-bench Verified, XS 2.1 scores 62.8% against XS.2's 59.9%. The model also leads the smaller-MoE category on SWE-Bench Pro and Terminal-Bench 2.0, though Poolside acknowledges the comparison includes dense models with larger activated parameter counts that trade efficiency for raw capability. This efficiency trend is part of a broader shift: teaching models to be better judges is now a key research direction.

What makes the numbers more credible is the methodology disclosure. Poolside ran all benchmarks using the Harbor Framework with 256K context, temperature 1.0, top_k 20, top_p 1, and thinking mode enabled. Each task ran in a sandbox with 8 GB RAM and 2 CPUs, except Terminal-Bench 2.0 which needed 48 GB RAM and 32 CPUs. The company also ran a reward-hack judge post-hoc on evaluation runs and reports finding no significant reward hacking.

What changes for local inference

The key operational news is that XS 2.1 ships with DFlash speculator draft models trained for each checkpoint. Poolside says these double the achieved tok/s in internal tests, making local inference noticeably faster. The model supports vLLM, SGLang, NVIDIA TensorRT-LLM, Hugging Face transformers, and Ollama out of the box, with llama.cpp support coming. Quantized checkpoints are available in FP8, INT4, and NVFP4 formats, which opens up deployment on hardware with tighter memory budgets. Speculative decoding is one of the smartest tricks for hiding latency, as covered in our deep dive on the technique.

This is where the release connects to a broader trend in local AI. A model that can do credibly well on coding benchmarks while keeping 3B active parameters is approaching the threshold where running a capable agent on a consumer GPU becomes practical, not just a demo. The DFlash speculators and quantized formats are the kind of infrastructure decisions that make a difference for actual users, not just benchmark hunters.

The license shift matters

Poolside is licensing XS 2.1 under OpenMDW-1.1, a permissive model-specific license that the company says is designed to reduce friction for open model distribution. The move aligns Poolside with the direction NVIDIA and the Linux Foundation are pushing with OpenMDW. It replaces the more restrictive terms that governed XS.2, which will sunset on Poolside's API in a week.

For developers considering whether to build on this model, the license change removes one barrier. OpenMDW-1.1 is designed for models and related artifacts and allows modification, use, and deployment without the kind of use-case restrictions that make some open-weight models legally uncertain for commercial projects.

Where the model sits in context

The comparison table Poolside published includes Qwen3.6 (35B-A3B), North Mini Code from Cohere (30B-A3B), and several larger models. XS 2.1 leads the smaller MoE models on most benchmarks shown. But the most interesting comparison may be with dense models. The table includes MAI-Code-1-Flash (137B-A5B) and GPT-5.4 Nano, both of which use more active parameters. XS 2.1 beats them on SWE-bench Multilingual and holds its own on other metrics, despite using far fewer active parameters.

This is the efficiency argument that MoE models have been making for years, and it's getting stronger. A 3B-active model that scores 63.1% on SWE-bench Multilingual would have been unthinkable two years ago. The question is whether benchmark gains translate to real coding productivity improvements, and that's harder to measure. For local deployment, the tradeoff between speed and privacy is crucial, as explored in our analysis of uncensored local models.

The practical takeaway

Poolside is pricing XS 2.1 at $0.10 per 1M input tokens, $0.20 per 1M output tokens, and $0.05 per 1M cache-read tokens, matching XS.2's pricing. The model is available through Poolside's API and OpenRouter. For local use, the combination of quantized weights, speculative decoding, and broad inference engine support means the gap between API-based coding and local coding is narrowing. This is reminiscent of how Cursor has redefined enterprise coding agents.

The sunset of XS.2 on Poolside's API after one week is aggressive but standard for model releases. The old model remains available through Baseten's Model Library for dedicated deployments. That gives teams who built on XS.2 a transition path.

None of this makes XS 2.1 the best coding model overall. Dense models with larger active parameter counts will still outperform it on many tasks. But the 3B-active MoE class is becoming a compelling option for scenarios where latency, privacy, or cost favor local inference. That is the trend this release feeds, and it is probably a more durable story than any single benchmark number.