Poolside Laguna: Open-Weight Coding Models That Punch Above Their Class

متوسط

On July 2, 2026, Poolside AI — a US-based foundation model lab that raised roughly $2B at a $12B valuation with NVIDIA participation — shipped its first public models: the Laguna family. Three MoE coding models, all released under the permissive OpenMDW-1.1 license, sharing a single pretraining corpus and differing mostly in how big a machine you point at them. Two weeks later they extended the family with Laguna S 2.1, a 118B/8B-active MoE that beats DeepSeek-V4-Pro-Max on DeepSWE v1.1 (40.4% vs 9.0%) while activating one-sixth the parameters, and lands 78.5% on SWE-Bench Multilingual — the top published score on that leaderboard at launch.

This page is the practical field guide for the family: what actually makes each size interesting, the two training-pipeline choices that most coverage skipped, the honest benchmark picture against closed frontier and open-weight competitors, how to run them locally today, and the specific tasks where reaching for a Laguna is the right call over Claude, GLM-5.2 or Kimi K3.

What you'll learn

• Know the family at a glance — three sizes, one corpus, one license — and which size maps to your hardware
• Understand why 8B active parameters can beat 1.6T total: MoE routing + coding-specialized post-training
• Read the honest benchmark picture across SWE-Bench Multilingual, Terminal-Bench 2.1, DeepSWE v1.1 and SWE-Bench Pro
• Know the two training-pipeline firsts that shape the model's behavior — RL in FP8 and multi-harness training
• Get Laguna running via OpenRouter in three lines, or locally via GGUF/MLX on a MacBook or DGX Spark
• Decide when Laguna is the right tool vs Claude Opus 5, GLM-5.2, Kimi K3 or GPT-5.5

The family at a glance

Three models, one shared pretraining run, three hardware envelopes. That is intentional: Poolside pretrained once and post-trained per size, so the personality is consistent across the family and you pick a size based on the box you have, not the task shape.

Model	Total params	Active params	Context	Target hardware	License
Laguna XS 2.1	33B (MoE)	3B	up to 256K	Single laptop / MacBook GPU	OpenMDW-1.1
Laguna S 2.1	118B (MoE)	8B	up to 1M	Single DGX Spark or H200 at FP8	OpenMDW-1.1
Laguna M.1	225B (MoE)	23B	256K+	Enterprise multi-GPU node	OpenMDW-1.1

The three practical takeaways from this table:

• Same corpus across the family. Laguna S 2.1's post-training added no new data on top of XS 2.1's — the S 2.1 announcement calls this out explicitly. Behavioral improvements come from the larger model class and a different post-training recipe, not from feeding it new documents. That means fine-tuning wisdom transfers cleanly across sizes.
• The active-param column is what matters for inference cost. At 8B active, Laguna S 2.1 has the compute profile of an 8B dense model per token, not a 118B one. That is the difference between "runs on a single accelerator" and "needs a rack".
• OpenMDW-1.1 is a real permissive license, not a source-available license. It was published by the Linux Foundation to close the gap between traditional open-source licenses (which don't cleanly cover weights and data) and vendor "open" licenses that quietly forbid commercial use. Grants unrestricted, royalty-free rights across copyright, patent, database and trade-secret; outputs are explicitly unencumbered.

آخر تحقّق: 2026-08-03.

Parameter counts, context windows and licensing are from Poolside's own launch materials and independent coverage. Benchmark scores in the next section are from Poolside's launch post and confirmed against MarkTechPost's analysis; leaderboard rank is quoted at launch and moves weekly. Confirm current numbers on Poolside's blog and the model card on Hugging Face before quoting them.

تأكّد من المصدر الرسمي.

Three things that aren't obvious from the headlines

1. RL in FP8 is a real engineering unlock, not marketing

Reinforcement learning for LLM post-training has historically run in BF16 because gradient noise at FP8 precision was assumed to blow up the tiny signal RL provides. Poolside publicly claims Laguna S 2.1 is the first frontier-scale model trained with RL in FP8. If true — and no counter-example has surfaced as of early August 2026 — this halves the memory footprint of the RL stage and lets a given cluster spend RL compute on tasks that used to be too expensive.

The practical consequence for users: Poolside's RL stage was selective. They report 409k training environments (83k terminal, 168k software-engineering), with SFT bootstrapping most behaviors from synthetic data and RL reserved for the highest-difficulty problems. That shows up in the benchmark deltas — the thinking-mode boost on DeepSWE v1.1 goes from 16.5% → 40.4% — which is the shape you'd expect when RL was spent narrowly on hard reasoning rather than smeared across every task.

2. Multi-harness training is why it doesn't fall over inside third-party agents

Coding models are usually trained inside one agent harness (the vendor's own scaffold) and then deployed inside dozens of others (Cursor, Cline, Aider, Continue, Kilo, etc.). The gap between training and deployment harness is where a lot of "the benchmark said it worked" failures come from. Poolside trained Laguna across 83k terminal environments and 168k software-engineering environments spanning multiple harnesses, which is the mechanism they credit for the model being usable outside their own scaffold on day one.

The visible tell of multi-harness training in Laguna's reported behavior: persistence (doesn't abandon an approach after one failed tool call), verification-first (re-reads its own output before shipping), and resourcefulness (finds alternate paths when the direct one is blocked). Those are agent behaviors, not next-token behaviors — you don't get them from more pretraining data.

The known limitations that map directly to this same design choice are worth noting so you're not surprised:

• Struggles with slight schema variations in third-party harnesses vs Poolside's native one — the harness has to look "close enough" to something in the 409k training envs.
• Sometimes emits incorrectly-escaped JSON in nested tool calls inside arrays. Rare but real.
• Overthinks math competition problems — extended thinking runs longer than needed on AIME-style tasks. Fine for coding, wasteful for math.

3. The active-parameter ratio is the whole story on cost

The eye-catching number is 1.6T vs 8B: Laguna S 2.1 beats DeepSeek-V4-Pro-Max (1.6T total) on DeepSWE v1.1 by 31.4 points while activating one-sixth the parameters per token. But the useful framing is the inference cost:

Model	Active params / token	Rough per-token compute	Reference API price*
Laguna XS 2.1	3B	~3B dense-equivalent	free-tier / $0.05 range
Laguna S 2.1	8B	~8B dense-equivalent	$0.10 in / $0.20 out per M (OpenRouter)
GLM-5.2	~40B	~40B dense-equivalent	$1.20-1.40 in / $4.10-4.40 out per M
Kimi K3	~32B	~32B dense-equivalent	similar to GLM range
Claude Opus 5 (frontier)	dense, undisclosed	frontier	$15 in / $75 out per M

*Prices from OpenRouter and equivalent aggregators, early August 2026 — treat as approximate.

The 150× per-token gap between S 2.1 on OpenRouter and Claude Opus 5 is not a fair fight for headline generation, but the useful ratio is S 2.1 vs the other open-weight frontier models (GLM-5.2, Kimi K3): Laguna is ~12× cheaper per output token, with SWE-Bench Multilingual scores in the same league. That is the price envelope that changes what work is economically sane to hand to a model.

Benchmark picture — the honest version

The scores that were reported at S 2.1's launch, cross-checked between Poolside's own post and third-party coverage:

Benchmark	Laguna S 2.1	What it measures	Notes
SWE-Bench Multilingual	78.5%	Real GitHub fix tasks across many languages	Top published open-weight score at launch
Terminal-Bench 2.1	70.2% (thinking) / 60.4% (non-thinking)	Long shell workflows	~10-point thinking-mode uplift
SWE-Bench Pro (Public)	59.4%	Harder autonomous repo-level fixes	Frontier closed models still ahead
DeepSWE v1.1	40.4% (thinking) / 16.5% (non-thinking)	Hard SWE benchmarks	Beats DeepSeek-V4-Pro-Max at 9.0% despite ~1/200 the active params
SWE Atlas (Codebase QnA)	46.2%	Understand large repos	Ahead of DeepSWE numbers
Toolathlon Verified	49.7%	Multi-tool orchestration	Mid-pack

Three interpretive rules for reading these:

• SWE-Bench Multilingual and Terminal-Bench 2.1 are the numbers to take seriously for an agentic coding model. They map to work you'd actually assign — "fix this bug in a real repo", "get this shell workflow to green". Poolside leads both on the open-weight side.
• Closed frontier still wins on the hardest benchmarks. Claude Fable 5 and Kimi K3 lead on SWE-Bench Pro and on some Toolathlon slices. Do not read Laguna's launch as "Anthropic is beaten" — read it as "the open-weight ceiling for coding just moved up several points".
• Thinking mode is not free but is where most of the win is. Terminal-Bench 60.4% → 70.2%, DeepSWE 16.5% → 40.4%. If you turn thinking off for cost reasons, you are giving up a large chunk of the model.

Getting Laguna running

Guided walkthrough1 of 4

1. OpenRouter, Baseten Model Library, Vercel AI Gateway and Kilo/Cline/Hermes integrations all offered day-one access. OpenRouter's list price at launch: $0.10 per M input, $0.20 per M output, with a free 256K-context tier for evaluation. OpenAI-compatible chat completions endpoint — swap the base URL and model ID.

Call Laguna S 2.1 via OpenRouter (Python)

import os, openai

client = openai.OpenAI(
  base_url="https://openrouter.ai/api/v1",
  api_key=os.environ["OPENROUTER_API_KEY"],
)

resp = client.chat.completions.create(
  model="poolside/laguna-s-2.1",
  messages=[
      {"role": "system", "content": "You are a senior engineer. Reply with a unified diff, no prose."},
      {"role": "user", "content": "Fix the off-by-one in this function:\n\ndef last_n(xs, n): return xs[-n-1:]"},
  ],
)
print(resp.choices[0].message.content)

When to reach for Laguna vs the alternatives

The pragmatic decision map, given the model landscape as of early August 2026:

Task	Best pick	Why
Long-horizon repo-level bug fix, budget-sensitive	Laguna S 2.1	Top open-weight SWE-Bench Multilingual, ~12× cheaper per output token than GLM-5.2 or Kimi K3
Agentic coding on a MacBook, offline required	Laguna XS 2.1	3B active params + MLX build; nothing else at this size hits comparable SWE-Bench Multilingual
Absolute frontier accuracy on the hardest tasks, cost no object	Claude Opus 5 or Fable 5	Closed frontier still leads SWE-Bench Pro and hardest slices
1M-token context on a repo, need permissive license	GLM-5.2 or Laguna S 2.1	Both offer 1M context under permissive licenses; GLM-5.2 is stronger on general reasoning, Laguna is stronger on agentic coding
Multi-tool orchestration where you need tight tool-call semantics	Claude or Kimi K3	Laguna's mid-pack Toolathlon score and known JSON-escaping edge case make it a weaker default here
Regulated environment, no data leaving the box	Laguna S 2.1 self-hosted	OpenMDW-1.1 grants unrestricted commercial use, US-origin lab; FP8 fits a single accelerator

The one-liner: Laguna is the current sweet spot when you want frontier-class coding behavior on a small active-parameter budget, and the license and origin matter. Use it when those constraints bind; reach for Claude or the closed frontier when they don't.

Related reading

• GLM-5.2: Open-Weight Frontier Coding Model — the direct comparison point on the other open-weight leader; different architectural trick (IndexShare) for the same 1M-context goal.
• Kimi K3 Open-Weight Frontier — the other major open-weight frontier coding model, denser params, different behavior envelope.
• DeepSeek & Qwen: Open Models Playbook — background on the open-weight coding landscape Laguna is competing with.
• Run Models Locally with Ollama — the local runtime you'll use for XS 2.1.
• Local Coding Agents — where XS 2.1 slots into the local-agent story.
• AI Gateways: LiteLLM, OpenRouter, Portkey — how to swap Laguna into an existing OpenAI-compatible stack.
• What AI Costs Across Providers — the price-per-task view that makes Laguna's ratio visible.

Check your grip

Check yourself

0/4

1. 1Laguna S 2.1 has 118B total parameters but only 8B active per token. Which model does it beat on DeepSWE v1.1 despite activating roughly one-sixth the parameters?
2. 2Why does the same 4,096-H200 pretrain get post-trained three different ways for XS 2.1, S 2.1, and M.1 — instead of just serving the biggest checkpoint everywhere?
3. 3You need to fix a bug in a 200-file repository and cost matters. Terminal-Bench 2.1 numbers matter. What's the most defensible pick among these, assuming you don't need frontier accuracy?
4. 4Which of these is a real, documented failure mode of Laguna S 2.1 that matters for production use?

Sources & further reading

• Introducing Laguna S 2.1 — Poolside — the launch post with training pipeline, benchmark table, and inference options.
• Poolside AI Launches Open-Weight 'Laguna' Coding Models — Open Source For You — family overview with XS 2.1 and M.1 specs.
• American AI startup Poolside launches free, high-performing open model Laguna XS.2 — VentureBeat — the July 2 launch coverage for the initial family.
• Poolside drops Laguna S 2.1, an open-weight coding model that beats rivals 10x its size — VentureBeat — post-launch analysis of S 2.1's comparative position.
• Poolside Releases Laguna S 2.1 — MarkTechPost — architecture and benchmark cross-check.
• Laguna XS 2.1 review — explainx.ai — hands-on impressions and hardware notes.
• OpenMDW license — Linux Foundation — the license under which all three Laguna models ship.