by Chris DePuy / July 28, 2026
This week, the local AI community has been running Poolside’s latest open-weight model, Laguna S 2.1, on everything from single DGX Sparks to 3090 Quad workstations. Released on July 21, Laguna S 2.1 is a 118B total parameter Mixture-of-Experts model with 8B activated parameters per token, a 1M-token context window, and both thinking and no-thinking inference modes. It went from the start of training to launch in under nine weeks, per Poolside’s blog post. We can confidently say in our own tests that Laguna runs very well on a single DGX Spark and can operate on a single prompt for hours at a time without getting lost, which is more than we can say for NVIDIA’s Nemotron Super (120B) or even DeepSeek V4 Flash running on two Sparks – Laguna is impressive.
Don’t just take it from us. Poolside and numerous others who have used the model recently have positive things to say.
Poolside’s own benchmarks show Laguna S 2.1 scoring 70.2% on Terminal-Bench 2.1 with thinking enabled — competitive with models many times its size, including 1T+ parameter models. The company positions it as “the most capable agentic coding model in its weight class by a wide margin.”
Local Inference Explosion
The community response has been immediate. The model’s open-weight release on Hugging Face and compatibility with standard inference engines triggered a wave of porting and benchmarking within days.
Alex Cheema noted that Laguna S 2.1 is “punching way above its weight class” — scoring higher than 1T+ parameter models on Terminal-Bench 2.1 while being small enough to run on a single DGX Spark. Sudo su open-sourced their full DGX Spark setup after spending a day resolving kernel flags, undocumented runtime dependencies, and memory wedge conditions — publishing the exact configuration needed to serve Laguna without hanging the box.
On the throughput side, Joe Muller (BosonJoe) benchmarked the model with DFlash speculative decoding on 2 DGX Sparks, finding that decode tok/s rises all the way to k=8 before tapering off, with an FP8 drafter and recipe on the way. Tech2Wild demonstrated Laguna S 2.1 on a 3090 Quad INT4 configuration hitting 256 tok/s peak and 249.9 tok/s sustained across four parallel coding agents with 0.12s average time-to-first-token — the fastest local model they have ever run.
Thinking Mode: A Dial, Not an Upgrade
Wësche published a detailed think ON vs OFF comparison on Laguna-S-2.1-NVFP4. The results show thinking mode improves code generation (70→73) and instruction following (86→92) but reduces agentic reliability from 98.6 to 86.3, with latency increasing from 5.9s to 7.8s. Overall TrueScore was 86.5 off versus 86.0 on. As Wësche concluded, “Thinking is a dial, not an upgrade. ON to write code. OFF to run agents.”
In a separate 74-task comparison, Wësche found that running Laguna S 2.1 locally (Int4-Int8Mix) scored 86.5 vs 80.9 running the same model in the cloud — part of a broader finding where local inference swept cloud 3-0 across GLM-5.2, Laguna S 2.1, and DeepSeek V4 on identical tasks and graders.
Refined for DGX Spark
The model was re-released on July 23 with NVFP4 quantization optimized for SM121 (the DGX Spark’s Grace-Blackwell GPU architecture). Community member mr-r0b0t updated the production contract to follow Poolside’s updated checkpoint while retaining vLLM 0.25.1, native SM121, FP8 KV cache, and a DFlash K=7 profile.
Broader Context
Laguna S 2.1 arrives in a period of rapid-fire open-weight releases. Kimi K3 (2.8T MoE, released July 27) brought native vision and a 1M context window. GLM-5.2 continues to see community porting to multi-DGX-Spark clusters. The common thread: model architects are designing for local deployment as seriously as for cloud inference, and the community is responding by publishing working recipes, kernel configurations, and benchmark suites for each release.
The model is available on Hugging Face with full benchmark trajectories published at trajectories.poolside.ai — open trajectories that Poolside describes as distinguishing genuine solutions from reward hacking via an LLM-as-a-Judge system calibrated on human-labeled trajectories.
Poolside has raised over $500 million to date, per its October 2024 announcement, and its commercial offering — the Poolside Platform — provides enterprise-grade agentic coding agents deployed inside customer security boundaries. The open-weight Laguna S 2.1 release serves as both a research contribution and a proof point for the platform’s underlying model capabilities.
For the local inference enthusiast, Laguna S 2.1 represents a significant data point: a 118B model that competes with 1T+ models on agentic coding benchmarks, runs on consumer-grade hardware, and has a growing ecosystem of community-tested recipes. As firms such as 650 Group have been tracking in their AI infrastructure research, the gap between frontier capability and local deployability continues to narrow.