On July 21, Poolside released a coding model that activates 8 billion parameters per token, scores 70.2% on Terminal-Bench 2.1, and fits on a single NVIDIA DGX Spark. The benchmark number is not the interesting part. The interesting part is what the MoE architecture does to your inference bill at scale - and what it means that a Western lab finally shipped an open-weight model into a size class that has been empty for nearly a year.
What Laguna S 2.1 actually is
Poolside spent most of its three-year existence quietly selling coding models to governments and defense agencies. Laguna S 2.1 is a 118-billion-parameter Mixture-of-Experts system that activates only 8 billion parameters per token and supports a context window of up to 1 million tokens. The weights shipped the same day on Hugging Face.
The active-parameter figure is what matters for teams thinking about cost. Laguna S 2.1's sparse MoE architecture - 256 routed experts plus one shared expert, with grouped-query attention and interleaved sliding-window layers - means inference costs scale with the 8 billion active parameters, not the 118 billion total. Put differently: inference cost is closer to an 8B model than a 118B model. You pay for 8B active parameters, not 118B.
Laguna S 2.1 shows capabilities in this domain that were reserved for frontier models just six months ago, at a small enough size to run on a single NVIDIA DGX Spark.
For the hardware-conscious: at 4-bit (NVFP4 or INT4), the weights need about 59 GB, which fits on a single DGX Spark's 128 GB of unified memory. At FP8 they need about 118 GB, still within a single Spark or a single H200. At BF16 they need about 236 GB, which calls for two linked Sparks or a multi-GPU datacenter node.
vLLM, SGLang, and Ollama support open serving and local inference from day one.
The Western open-weight gap this fills
Laguna S 2.1 occupies a size class that no Western lab has released an open-weight model into in 11 months, since gpt-oss-120b's release in August last year.
That gap is the strategic context for this release - not any individual benchmark score.
The trajectory is unusual: the April dual release of Laguna M.1 and XS.2, the July 2 refresh of XS 2.1, and now S 2.1, which Poolside says outperforms April's flagship M.1 at roughly a third of its active size.
Remarkably, S 2.1 used the exact same pre-training data as XS 2.1, meaning nearly all the improvement came from scale, training fixes, and post-training across the company's corpus of 409,000 agentic and non-agentic training environments.
Poolside's training approach is genuinely novel: RLCEF (Reinforcement Learning with Code Execution Feedback) means instead of human preferences, the model learns from actual code execution results - across 409,000 training environments that are real, diverse software engineering scenarios, using multi-harness rollouts where the model attempts problems multiple ways and learns from all outcomes.
How the pricing actually stacks up
This is where teams doing real workload math need to sit down with the actual numbers rather than headline comparisons.
Paid API on OpenRouter runs $0.10/M input and $0.20/M output tokens with the full 1M context. For comparison, Claude Sonnet 5 costs $2/$10, and Grok 4.5 costs $2/$6. Laguna S 2.1 is 20× cheaper than Sonnet on input and 50× cheaper on output.
Against GPT-5.6 Sol specifically: Sol prices at $5 input and $30 output per million tokens. That makes Laguna S 2.1 50× cheaper on input and 150× cheaper on output at list prices.
| Model | Input $/M | Output $/M | Terminal-Bench 2.1 | Weights available |
|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | 88.8% | No |
| GPT-5.6 Luna | $1.00 | $6.00 | Not published | No |
| Laguna S 2.1 (API) | $0.10 | $0.20 | 70.2% | Yes |
| Laguna S 2.1 (self-hosted) | ~$0 marginal | ~$0 marginal | 70.2% | Yes |
| GLM-5.2 (OpenRouter) | $0.95 | $3.00 | 91.2% GPQA | Yes |
The caveat the price table obscures: the frontier remains clearly out of reach - closed models like GPT-5.6 Sol at 88.8% on Terminal-Bench 2.1 and Claude Fable 5 at 88.0%, along with the 2.8-trillion-parameter open-weight Kimi K3 at 88.3%, sit well above Laguna S 2.1. An 18-point gap on a coding benchmark is not nothing. For a team running a coding agent that touches production deployments or generates security-sensitive code, the gap is worth the cost. For teams running code review, documentation generation, test scaffolding, or internal tooling, 70.2% may be more than sufficient at a fraction of the price.
There is also a harness-consistency caveat worth noting: Poolside's methodology takes the maximum of vendor self-reported scores, benchmark-author leaderboards, and third-party figures for competitors - a reasonable convention, but one that mixes harnesses and test conditions. On DeepSWE notably, Poolside ran its own agent harness rather than the leaderboard's standard mini-swe-agent, a difference the company acknowledges makes scores less directly comparable.
What this means for teams deciding on a coding model
For privacy-sensitive teams or organizations in regulated industries, the self-hosting story is the real news. For government, defense, and other highly regulated organizations that Poolside already serves, self-hosting keeps sensitive code and data inside their own environment. With Laguna S 2.1 fitting on a single DGX Spark, that option is now within reach for teams that cannot afford a multi-node GPU cluster.
For cost-sensitive teams with high coding-agent volume, the economics shift materially. A team running an agent that generates 10 million output tokens a day - plausible for a mid-size engineering org using Cline, OpenCode, or a similar harness - pays $200/day at Laguna S 2.1 API rates versus $3,000/day at GPT-5.6 Sol rates. Self-hosted eliminates the variable cost entirely, replacing it with a fixed hardware line item.
A practical routing strategy: use Sol or an equivalent frontier model for critical production code generation where correctness failures are expensive, route documentation, test generation, code review, and internal tools through Laguna S 2.1 or another cost-efficient open model, and re-run your evals quarterly as the open-weight tier catches up.
Beagle in action#engineering-infra, 2:43pmPoolside says its next, larger Laguna model began pre-training last week. The cadence - Poolside ships on a roughly five-week cadence
- suggests the gap between open-weight and frontier models in this category will keep narrowing faster than the annual model cycles most teams plan around.
Open-weight agentic coding models: common questions
What is Laguna S 2.1 and who made it?
Laguna S 2.1 is a 118-billion-parameter Mixture-of-Experts coding model released by Poolside on July 21, 2026. It activates 8 billion parameters per token, supports a 1-million-token context window, and is designed specifically for agentic coding tasks - multi-step, tool-using workflows rather than isolated code generation.
Can you run Laguna S 2.1 locally without a GPU cluster?
Yes, at 4-bit quantization the weights require about 59 GB of memory, which fits on a single NVIDIA DGX Spark with its 128 GB of unified memory. The model runs via Ollama, vLLM, or SGLang from day one. At BF16, you need around 236 GB - two linked Sparks or a multi-GPU node.
How does open-weight agentic coding model pricing compare to GPT-5.6?
At OpenRouter list prices, Laguna S 2.1 costs $0.10/M input and $0.20/M output tokens. GPT-5.6 Sol costs $5/M input and $30/M output - a 50× and 150× gap respectively. Self-hosted Laguna S 2.1 has near-zero marginal cost. The tradeoff is an 18-point gap on Terminal-Bench 2.1.
Does the open-weight license allow commercial use?
Laguna S 2.1 is released under the OpenMDW-1.1 license. Poolside's primary target is government and regulated enterprise customers, where self-hosted open weights are a requirement. Check the specific license terms on the Hugging Face model card before production deployment.
When should a team still use a frontier closed model over Laguna S 2.1?
When correctness failure is expensive: production deployment code, security-sensitive refactors, or anything where a wrong answer causes a customer-facing incident. The 18-point Terminal-Bench gap between Laguna S 2.1 and GPT-5.6 Sol is real. For test generation, documentation, code review summaries, and internal tooling, run your own evals - the cost gap may justify routing those tasks to the open-weight tier.