Poolside AI Open-Weights Laguna: Self-Hosted vs API Costs for Coding Teams
By Eric Bush · July 9, 2026 · 8 min read
Open Weights Change the Calculation
Poolside AI has released their Laguna coding models as open-weight as of July 8–9, 2026. The flagship Laguna XS 2.1 — a 33B parameter model with 3B active parameters (MoE architecture) and 262K context window — was already available via OpenRouter at $0.06 input / $0.12 output per million tokens. Now with open weights, teams can self-host it.
This creates a genuine decision point that did not exist before: should your team pay per-token via API, or invest in GPU infrastructure to run Laguna locally? The answer depends entirely on your monthly token volume, and the math is more nuanced than most teams realize.
Self-Hosting Hardware Requirements
Laguna XS 2.1 has 33B total parameters with only 3B active per forward pass (thanks to MoE). However, you still need to load all 33B parameters into GPU memory. In FP16, that requires approximately 66GB of VRAM — which fits on a single NVIDIA A100 80GB or H100 80GB.
With INT8 quantization, the memory requirement drops to ~33GB, fitting comfortably on a single A100 40GB. With INT4 quantization (~17GB), you could potentially run it on consumer hardware like an RTX 4090 (24GB), though with some quality degradation.
Cloud GPU costs: A100 80GB instances run approximately $2/hour on major cloud providers. H100 instances cost $3–3.50/hour. Reserved instances with 1-year commitments can reduce this by 30–40%.
Cost Per Million Tokens: Self-Hosted vs API
Let us calculate the self-hosted cost. A 33B MoE model with 3B active parameters on an A100 achieves approximately 150–200 tokens/second for generation (output). That means one A100 produces roughly 540K–720K tokens per hour.
At $2/hour for an A100 and ~630K tokens/hour average output: $2 / 0.63M = ~$3.17 per million output tokens. For input (prefill), throughput is much higher — roughly 2,000–3,000 tokens/second — giving approximately $0.20–0.28 per million input tokens.
Wait — that is significantly more expensive than the API. The OpenRouter API charges $0.06/$0.12, while self-hosting costs ~$0.25/$3.17. How is the API so much cheaper?
The answer is batch density. API providers run dozens of concurrent requests on the same GPU, amortizing the hardware cost across many users. A single team self-hosting dedicated hardware cannot achieve the same utilization unless they maintain consistently high request volume.
The Break-Even Analysis
Self-hosting becomes competitive only when you can maintain near-100% GPU utilization. Let us model the break-even:
Monthly A100 cost (on-demand): $2/hr × 730 hrs = $1,460/month. Monthly A100 cost (1-year reserved): ~$1,000/month. At the API rate of $0.12 per million output tokens, you would need to generate $1,460 / $0.12 = 12.2 billion output tokens/month via API before self-hosting saves money — assuming 100% utilization.
That is an absurd volume for a single team. At realistic 40–60% utilization (accounting for off-hours, variable demand), the break-even pushes even higher. For most teams under 50 developers, the API at $0.06/$0.12 is dramatically cheaper than self-hosting.
When Self-Hosting Actually Wins
Self-hosting Laguna makes economic sense in specific scenarios that have nothing to do with per-token cost:
Data privacy requirements: If your code cannot leave your infrastructure (defense, healthcare, regulated industries), self-hosting is not about cost — it is about compliance. The $1,460/month is your compliance premium.
Latency sensitivity: Self-hosted models eliminate network round-trip time. For IDE autocomplete where 50ms matters, local inference can provide a noticeably better developer experience.
Rate limit freedom: No API throttling, no quota caps, no waitlists. If you have bursty workloads that hit API rate limits (CI/CD pipelines running hundreds of parallel agent tasks), owned hardware removes that constraint.
Comparison to Other Open Coding Models
Laguna XS 2.1 enters a crowded field of open-weight coding models. DeepSeek Coder V3 (236B MoE, 37B active) requires 2–4 A100s to self-host but offers stronger benchmark scores. CodeLlama 34B is similar in size but older and less capable. Qwen 3.6 35B offers comparable performance with broader general capabilities.
Laguna's advantage is the 262K context window and its purpose-built coding focus. For teams that need long-context code understanding (large monorepos, long file refactoring), Laguna may outperform larger models that cap at 32K–128K context.
Recommended Strategy by Team Size
Solo developers and small teams (1–10): Use the API at $0.06/$0.12. It is absurdly cheap and you will never reach volumes that justify infrastructure management overhead.
Mid-size teams (10–50): Use the API as primary, but consider self-hosting for specific workloads like CI/CD pipelines or internal tools where you can guarantee consistent utilization.
Enterprise teams (50+) with compliance needs: Self-host on reserved instances. At this scale you can batch enough requests to improve utilization, and the compliance benefit alone justifies the premium. Budget $1,000–3,000/month per GPU node depending on your reserved pricing.
Want to calculate exact costs for your project?
Frequently Asked Questions
How much does it cost to self-host Poolside Laguna XS 2.1?
Self-hosting on a single A100 80GB costs approximately $1,460/month on-demand or $1,000/month with reserved pricing. The model fits on one GPU thanks to its MoE architecture (33B total, 3B active parameters).
Is self-hosting Laguna cheaper than using the API?
For most teams, no. The API at $0.06/$0.12 per million tokens is dramatically cheaper because providers achieve high GPU utilization through batching. Self-hosting only makes sense for compliance requirements, latency needs, or rate limit freedom — not cost savings.
What GPU do I need to run Poolside Laguna locally?
In FP16, you need 66GB VRAM (A100 80GB or H100). With INT8 quantization (~33GB), an A100 40GB works. With INT4 quantization (~17GB), even an RTX 4090 24GB can run it, though with some quality loss.
How does Laguna compare to DeepSeek Coder V3 for self-hosting?
DeepSeek Coder V3 (236B MoE, 37B active) requires 2-4 A100s and offers stronger benchmarks but higher infrastructure cost. Laguna XS 2.1 fits on a single GPU and has a 262K context advantage, making it more practical for smaller self-hosted deployments.
At what volume does self-hosting Laguna break even vs the API?
At the API rate of $0.12 per million output tokens, you would need approximately 12.2 billion output tokens per month to justify a single dedicated A100. This is unrealistic for teams under 50 developers, making the API the better choice for most.
Related Articles
Self-Hosted vs API AI Coding: Total Cost of Ownership in 2026
A comprehensive TCO analysis comparing self-hosted open-source models against cloud API services for AI coding in 2026. Covers hardware costs, operational overhead, and the crossover points where each approach wins.
Open-Weight vs API: The True Cost of Running Coding Models on Your Own GPU
Compare the real costs of self-hosting open-weight coding models like DeepSeek Coder and Qwen2.5-Coder on GPU versus using API providers. Includes break-even analysis and recommendation matrix.
Poolside Laguna XS 2.1 at $0.06/$0.12: A New Ultra-Cheap Coding Model on OpenRouter
Poolside Laguna XS 2.1 appeared on OpenRouter on July 2, 2026 with a 262,144-token context window and $0.06/$0.12 per million token pricing. Here is how to evaluate it for low-cost agentic coding.