Poolside Laguna 2.1 - 33B and 118B open-weight coding models under OpenMDW-1.1

Poolside put two agentic coding models out as open weights inside three weeks, both under OpenMDW-1.1 - the Linux Foundation license that covers weights, code and data as one unit rather than bolting Apache onto the repo and leaving the weights undefined.

The non-obvious part: S 2.1 was trained on the exact same pre-training data as XS 2.1. Poolside's own framing is that nearly all the gain came from scale, training fixes and post-training - which says the data pipeline was never the bottleneck.

The two models

	Laguna XS 2.1	Laguna S 2.1
Released	2 July 2026	21 July 2026
Params	33B total, 3B active	118B total, 8B active
Context	256K	1M
License	OpenMDW-1.1	OpenMDW-1.1
Formats	BF16, FP8, INT4, NVFP4, GGUF	BF16, FP8, INT4, NVFP4, GGUF, MLX

Benchmarks, honestly read

Poolside pitches S 2.1 as the West’s most capable open-weight model. On the numbers it is competitive in its class and clearly behind the frontier:

Benchmark	Laguna S 2.1	Best listed competitor
SWE-Bench Multilingual	78.5%	Qwen 3.7 Max 78.3%
Terminal-Bench 2.1	70.2%	Kimi K3 88.3%
SWE-Bench Pro	59.4%	Claude Fable 5 80.3%
DeepSWE v1.1	40.4%	GPT-5.6 Sol 73.0%

SWE-Bench Multilingual is the one it wins, against models several times its size. Terminal-Bench and DeepSWE it does not. Thinking mode is worth a lot here - Terminal-Bench 2.1 goes from 60.4% to 70.2% with it on.

XS 2.1 scores 63.1% on SWE-Bench Multilingual, up 5.4 points on XS.2, evaluated at 500 max steps with temperature=1.0, top_k=20, top_p=1 and thinking enabled.

What actually fits

Laguna-XS-2.1-Q4_K_M.gguf is 20.3 GB. That clears a 24GB 4090 with room for a modest context, and sits comfortably on a 32GB 5090. BF16 is 66.9 GB.

The S quant ladder from unsloth tells a different story:

Quant	Size
UD-IQ1_S	33.8 GB
UD-Q2_K_XL	39.7 GB
UD-Q3_K_XL	54.1 GB
UD-IQ4_XS	57.6 GB
UD-Q4_K_XL	73.4 GB
Q8_0	125.0 GB
BF16	235.2 GB

The smallest quant on offer is 33.8 GB, and a 1-bit quant of a coding model is not a serious proposal. S 2.1 is a two-card or CPU-offload model on consumer hardware. Attention layout is 12 global attention layers plus 36 sliding-window layers at window 512, and the GGUF carries the full 1,048,576-token context.

Unsloth’s llama.cpp invocation, with --fit doing the offload split:

./llama.cpp/build/bin/llama-server \
  --model Laguna-S-2.1-GGUF/UD-Q4_K_XL/Laguna-S-2.1-UD-Q4_K_XL-00001-of-00003.gguf \
  --fit on --ctx-size 16384 --port 8000

Thinking is per-request, and preserved thinking is the recommendation for agentic coding:

extra_body={"chat_template_kwargs": {"enable_thinking": False}}

The licensing angle is the real news. Laguna is why license:openmdw-1.1 exists as a Hugging Face facet at all - 145 model repos carry the tag today and most of them are Laguna and its quantizations. NVIDIA has committed to the same license for future Cosmos, Isaac GR00T, Ising and Nemotron releases. See the license roundup for what OpenMDW-1.1 actually grants.