Laguna: From a 33B Local Workhorse to 118B Long-Horizon Reasoning, Poolside's Three-Releases-in-Three-Months Bet

🌏 中文版

On July 21, 2026, Poolside released Laguna S 2.1 — a 118B-total, 8B-active MoE with 1M context, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE. It went from the start of training to the launch post in under nine weeks, on exactly the same pre-training data as the small model from two months earlier. This is the twenty-first family deep-dive in the "AI 模型家族" series, tracing Laguna from the M.1 / XS.2 dual release through XS 2.1 to S 2.1.

For how to read benchmark numbers in this post, see the AI model evaluation sources guide. This post is part of the AI model landscape overview series.

Family timeline

Three models in three months. The arc is clear: use M.1 / XS.2 to expose weaknesses in the harness, data, and training loop, validate the improved recipe on XS 2.1, then scale the whole thing to S 2.1. The S 2.1 launch post puts it plainly: "under nine weeks from the start of training to this post", powered by the internal Model Factory platform. The next, larger Laguna had already started pre-training a week before the post.

Two sizes, one philosophy

To understand Laguna, accept its default first: these models are not for chatting, they are for running hundreds of steps inside an agent harness. The docs say the same thing about M.1 and XS.2: use them inside Poolside Agent workflows, not as standalone chat models. S 2.1 and XS 2.1 are no different — supported modes are Agentic + Chat, but peak results all come from the in-house pool terminal coding agent.

XS line (runs locally): XS 2.1 is 33B total with 3B active and 256K context (262,144 exactly). The official line is "compact enough to run on a Mac with 36 GB of RAM", with FP8, NVFP4, and INT4 checkpoints on Hugging Face plus a DFlash draft model that doubles throughput in testing. Paid pricing carries over from XS.2 at $0.10 / $0.20 / $0.05 per 1M input / output / cache-read tokens.

S line (holds the long run): S 2.1 is 118B total with 8B active and 1M context (1,048,576 exactly). The pitch is "Frontier-class reasoning at mid-size cost", small enough for a single NVIDIA DGX Spark yet strong enough to beat several 500B+ open models on Terminal-Bench 2.1. On OpenRouter, the free endpoint offers 256K context while the dedicated paid endpoint unlocks the full 1M at $0.10 / $0.20 / $0.01 (announced July 2026).

Neither size supports vision — text in, text out only. That is deliberate, not an omission: Laguna bets all complexity on the read-repo, edit-file, run-test, verify loop.

Architecture: one laguna recipe

S 2.1 and XS 2.1 share the same laguna architecture and differ only in scale. Per the S 2.1 model card and the XS 2.1 model card:

• Routing: token-choice router with softplus gating, 256 routed experts plus 1 shared expert; S 2.1 takes top-10 per token for about 8B active
• Attention: grouped-query attention with 8 KV heads, head dim 128, plus per-head softplus output gating
• Layers: S 2.1 has 48 layers (12 global + 36 sliding-window), XS 2.1 has 40 layers (10 global + 30 sliding-window), both at a 1:3 ratio with a 512-token sliding window
• Vocabulary: shared family tokenizer at 100,352 tokens
• Reasoning: native interleaved thinking that reasons between tool calls; keeping prior reasoning blocks in history works best

S 2.1's BF16 weights run about 236GB and need multiple GPUs; quantized variants lower the bar. XS 2.1 quantizes the KV cache to FP8 so single-GPU and Mac setups stay viable. The standard MoE trade-off applies: total parameters set the memory floor, active parameters set the per-token cost.

Training: same data, different post-training

The most interesting line in the S 2.1 post is this: "trained on exactly the same pre-training data as XS 2.1". The step from XS 2.1 to S 2.1 was scale, training-code fixes, and small recipe changes — not new data. S 2.1 pre-training started May 22, 2026 on 4,096 NVIDIA H200 GPUs for 60 days, listed on the official site at 30T tokens (XS 2.1 is listed at 15T). It is also Poolside's first model with RL done entirely in FP8 precision.

Most of the gap comes from post-training, in two stages. First SFT bootstraps capabilities with partly synthetic data, then RL is reserved for tasks the model still cannot solve at a high pass rate. The corpus spans 409K agentic and non-agentic environments, with 83K terminal setups and 168K standard software-engineering flows. The largest software-engineering source reproduces real commits (about 38,000 tasks across about 17,000 repositories), and S 2.1 adds agentic repository installation: given a repo, install every dependency until the test suite runs.

Three training-loop changes map directly to the "keeps going" behavior: more generous rollout budgets (longer timeouts, more tokens per turn, more turns per task), a new sandbox service (background processes, selective network blocking against reward hacking), and multi-harness rollouts (same prompts across several harnesses to avoid overfitting to one). Quoting co-head of applied research Pengming Wang in the post: "not necessarily add more intelligence, but improve the behaviors" — more verification, fewer assumptions, no early victory laps, more persistence.

The thinking toggle is the receipt: S 2.1 ships with only off and max modes (max by default), and max lifts Terminal-Bench 2.1 from 60.4% to 70.2% and DeepSWE from 16.5% to 40.4%. The cost is mean completion tokens rising from 80K to 129K (TB 2.1) and 99K to 249K (DeepSWE). The disclosed limitations are refreshingly blunt: third-party harnesses with slightly different tool schemas can trigger memorized interfaces, JSON-array arguments can break escaping, and math problems can overthink.

Family matrix and selection

The selection logic is blunt: for local, fast, and cheap — XS 2.1; for 1M long runs and leaderboard scores — S 2.1. M.1 and XS.2 are transitional now; XS.2 sunset on the in-house API a week after XS 2.1 launched and lives on only as Baseten dedicated deployments.

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("POOLSIDE_API_KEY"),
    base_url="https://inference.poolside.ai/v1",
)
resp = client.chat.completions.create(
    model="poolside/laguna-s-2.1",
    messages=[{"role": "user", "content": "What are channels in Go?"}],
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(resp.choices[0].message.content)

Open strategy and ecosystem

Both current Laguna models ship under OpenMDW-1.1, fully permissive with no extra permission needed for commercial use. S 2.1 arrived day one with BF16, FP8, INT4, and NVFP4 plus official GGUF and MLX conversions; XS 2.1 offers BF16, FP8, NVFP4, and INT4. The M.1 era used Apache 2.0, and the 2.1 generation moved wholesale to OpenMDW, following the NVIDIA and Linux Foundation direction.

The serving ecosystem was day-one complete: vLLM, SGLang, Ollama (ollama run laguna-s-2.1), llama.cpp, MLX, NVIDIA TensorRT-LLM, plus Baseten, OpenRouter, and Vercel AI Gateway. On the agent side it landed in Kilo, Hermes Agent, pi, OpenCode, OpenClaw, and Cline. Every published score ships with full trajectories on trajectories.poolside.ai, which is more accountability than most score-only releases.

Where it stands against rivals

Using the official July 21, 2026 table, S 2.1 is best described as "best fighter outside its weight class":

• Terminal-Bench 2.1 at 70.2% beats Inkling (975B-A41B, 63.8%), DeepSeek-V4-Pro-Max (1.6T-A49B, 64.0%), and Nemotron 3 Ultra (550B-A55B, 56.4%), but trails Tencent Hy3 (295B-A21B, 71.7%) — let alone flagships like Kimi K3 (2.8T, 88.3%) and Claude Fable 5 (88.0%).
• DeepSWE at 40.4% sits with GLM 5.2 (44.0%) and far below Kimi K3 (69.0%), Claude Fable 5 (70.0%), and Muse Spark 1.1 (53.3%). But it leaves its own XS 2.1 (0.3%) and DeepSeek-V4-Pro-Max (9.0%) far behind — the long-horizon gap is stark.
• SWE-bench Multilingual at 78.5% is actually a bright spot, ahead of Hy3 (75.8%) and Qwen 3.7 Max (78.3%), with SWE-Bench Pro at 59.4% only slightly behind Qwen (60.6%) and Muse Spark (61.5%).

XS 2.1 fights at the 30B level: SWE-bench Verified at 70.9% beats Cohere North Mini Code (67.6%) but trails Qwen3.6-35B-A3B (73.4%) and Claude Haiku 4.5 (73.3%); Terminal-Bench 2.0 at 37.5% beats Haiku (29.8%) but trails Qwen (51.5%) and GPT-5.4 Nano (46.3%). Same conclusion as the official one: leads its weight class, but punching up has a ceiling.

What it means for agent builders

For long-horizon coding agents (multi-file edits, test runs, self-verification): S 2.1 is one of the few open models explicitly tuned for "hundreds of steps without giving up". Keep thinking on, preserve reasoning, and run it in pool or a compatible harness — otherwise DeepSWE-style tasks fall from 40.4% back to 16.5%.

For local or low-cost agents (laptop demos, offline setups, fast iteration): XS 2.1 is the saner starting point. It runs on a 36GB Mac, installs via one Ollama command, doubles throughput with DFlash, and its 63.1% Multilingual score is solid at 33B.

Not for: vision (the whole family is text-only), adjustable effort (S 2.1 offers only off / max, no low / mid / high), or zero-friction third-party harnesses (known schema-memory quirks need harness rejections plus retries).

Overall

Laguna's bet is not "bigger pre-training" but "better working behavior" — verify, backtrack, persist. S 2.1 shows the same data with different post-training and a different training loop can jump from XS-level to challenging 500B+ leaderboard entries. The next step is already announced: a larger Laguna is in pre-training, aiming to scale both size and that persistence together.

Two things to watch: how many steps S 2.1's 1M context actually holds together in real repos outside benchmark sandboxes, and whether the OpenMDW-1.1 open strategy buys enough third-party harness adoption to fix the known tool-call rough edges.

References

• Poolside: Introducing Laguna S 2.1 (2026-07-21)
• Poolside: Introducing Laguna XS 2.1 (2026-07-02)
• Poolside: Models (S 2.1 / XS 2.1 specs and pricing)
• Poolside docs: Model release notes (M.1 / XS.2 / XS 2.1 / S 2.1)
• Hugging Face: poolside/Laguna-S-2.1
• Hugging Face: poolside/Laguna-XS-2.1
• OpenRouter: poolside/laguna-s-2.1
• trajectories.poolside.ai: full S 2.1 evaluation trajectories