What Is Laguna S 2.1? The 118B-A8B Open MoE Scoring 70.2% on Terminal-Bench

Laguna S 2.1 is an open-weight Mixture-of-Experts coding model released by Poolside on July 21, 2026, with 118 billion total parameters, 8 billion activated per token, and a context window up to 1 million tokens. It scores 70.2% on Terminal-Bench 2.1 — placing 11th on Poolside's compiled leaderboard, ahead of open models several times its size — and its NVFP4 weights fit on a single NVIDIA DGX Spark. The weights are on Hugging Face under the OpenMDW-1.1 license, and the model is available today via OpenRouter, Ollama, vLLM, and Poolside's own agent CLI.

Key Takeaways

• 118B total / 8B active Mixture-of-Experts, 256 routed experts (top-10 per token) plus 1 shared expert, 48 layers.
• 1M-token context on the paid endpoint; 256K on the free tier and local Ollama builds.
• 70.2% on Terminal-Bench 2.1 using Poolside's own pool harness with thinking enabled.
• Open weights under OpenMDW-1.1 — commercial use permitted, weights on Hugging Face day one.
• Runs on a single 128GB machine — Apple Silicon or NVIDIA DGX Spark — thanks to the small active-parameter count.
• Trained in under nine weeks on 4,096 NVIDIA H200 GPUs; pretraining started May 22, 2026.

What Is Laguna S 2.1?

Laguna S 2.1 is Poolside's most capable open-weight model to date, purpose-built for agentic coding and long-horizon software engineering work. It sits in the "S" size class of Poolside's Laguna family — larger than Laguna XS 2.1 (33B) but far smaller than the trillion-parameter open MoEs it competes with on benchmark leaderboards.

The model was announced on July 21, 2026 in a Poolside blog post and simultaneously published on Hugging Face at poolside/Laguna-S-2.1. Poolside co-CEO Jason Warner framed the release as a response to the shortage of Western open-weight models in this size class: "The West needs open-weight models it can trust, run, and build on." The company positions Laguna S 2.1 as the first open-weight model in its size band to come out of a Western lab in 11 months, since OpenAI's gpt-oss-120b in August 2025.

Laguna S 2.1 supports both a thinking mode and a no-thinking mode, controllable per request, and its 1M-token context window is designed for extended reasoning traces and long agent trajectories.

Laguna S 2.1 at a Glance

Spec	Value
Developer	Poolside
Type	Mixture-of-Experts, dense-attention transformer
Total parameters	118B
Active parameters per token	8B
Experts	256 routed (top-10) + 1 shared
Layers	48 (12 global attention + 36 sliding-window at 512)
Attention	Grouped-query attention, 8 KV heads
Context window	1,048,576 tokens (paid); 262,144 tokens (free tier / Ollama)
Max output	131,072 tokens (paid); 32,768 (free)
Modality	Text in, text out
Tools	Function calling, interleaved thinking
License	OpenMDW-1.1 (commercial use permitted)
Release date	July 21, 2026
Pricing (OpenRouter paid)	$0.10 / M input · $0.20 / M output · $0.01 / M cache-read
Pricing (free tier)	$0.00
Availability	Hugging Face weights, OpenRouter, Ollama, Vercel AI Gateway, Poolside API, OpenCode

How Laguna S 2.1 Works

The architecture is a sparse Mixture-of-Experts with a distinctive attention pattern. Poolside alternates one global attention layer with three sliding-window attention layers set to 512 tokens, giving the model long-context reach without paying the full quadratic cost. Grouped-query attention with 8 KV heads keeps the KV cache small at 1M-token context.

The expert routing uses sigmoid gating with normalized expert weights and per-layer rotary scales, rather than the more common softmax top-k. 256 routed experts plus one shared expert are active in each MoE block, with the top 10 routed experts firing per token.

The most distinctive runtime feature is DFlash, a paired speculative-decoding draft model (poolside/Laguna-S-2.1-DFlash-NVFP4) that ships alongside the main weights. In Poolside's own tests, DFlash pushes single-stream decode from around 109 tok/s to 271 tok/s on code prompts on an RTX 6000 PRO. It is less useful on prose and can be a net loss under high concurrency, since rejected drafts eat batch capacity — a caveat community testers have flagged.

Poolside also published full evaluation trajectories for every benchmark trial at trajectories.poolside.ai, a transparency move meant to preempt "benchmaxxing" skepticism. Every score is pass@1 averaged over 3–4 attempts per task.

What You Can Do With Laguna S 2.1

The model is optimized for agentic coding and long-horizon software tasks. Concrete workloads seen in the first day of community use:

• Multi-step tool chains. One independent evaluator reported successful 6-level tool chains in their private harness, the deepest they had measured on any local model.
• Codebase question-answering at long context, backed by the SWE Atlas score of 46.2%.
• Terminal-based agent work — resolving tasks in a shell, editing files, running commands — the workload Terminal-Bench 2.1 measures.
• Local coding copilots. Thomas Wolf of Hugging Face called it "quite possibly the best coding model you can run locally (single DGX spark or Mac) atm".
• Data-driven app scaffolding. The Kilo Code team fed it NYC Open Data on release day and had it build a topographical borough visualizer as a first test.

Ivan Fioravanti published early throughput numbers on Apple M5 Max with Ollama, showing 49 tok/s decode at 0.5K context, dropping to 36 tok/s at 16K.

Source: @ivanfioravanti

How Laguna S 2.1 Compares

Poolside publishes a compiled leaderboard covering the top agentic coding models. Key comparisons:

Model	Total / Active	Terminal-Bench 2.1	SWE-Bench Multilingual	SWE-Bench Pro (public)
Laguna S 2.1	118B / 8B	70.2%	78.5%	59.4%
Tencent Hy3	295B / 21B	71.7%	75.8%	57.9%
Inkling	975B / 41B	63.8%	—	54.3%
DeepSeek-V4-Pro-Max	1.6T / 49B	64.0%	76.2%	55.4%
Kimi K3	2.8T / 50B	88.3%	—	—
Claude Fable 5	undisclosed	88.0%	—	80.3%
GPT-5.6 Sol	undisclosed	88.8%	—	—

Laguna S 2.1 trails closed frontier models like Claude Fable 5 and GPT-5.6 Sol on Terminal-Bench, but ranks first among models with 8B or fewer active parameters. Against similar-scale open-weight peers like Hy3 and Inkling, it is within a point on Terminal-Bench and slightly ahead on SWE-Bench Pro despite its much smaller active-parameter count. Our comparison of Kimi K3 vs Claude covers how the closed frontier stacks up on similar workloads.

Availability: How to Access Laguna S 2.1

• Hugging Face — poolside/Laguna-S-2.1 (BF16), plus Laguna-S-2.1-NVFP4, Laguna-S-2.1-DFlash-NVFP4, and Laguna-S-2.1-GGUF for llama.cpp.
• Ollama — ollama run laguna-s-2.1, with tags latest, q4_k_m (75GB), q8_0 (128GB), and f16 (235GB), all at 256K context.
• vLLM and SGLang — day-one native support; vllm serve "poolside/Laguna-S-2.1".
• OpenRouter — free tier at 262K context and a paid endpoint at 1M context.
• Poolside Agent CLI — the pool command, documented at docs.poolside.ai.
• Vercel AI Gateway, Baseten, OpenCode, Kilo Code, Cline — all wired up on release day.

If you want to run a comparable frontier coding chat model through a hosted API, Kimi K3 is available on kie.ai and sits at the top of the Terminal-Bench 2.1 leaderboard among open-weight models.

Community benchmarks confirm the "single-machine" claim is real but tight. On a single NVIDIA DGX Spark, decode sits at 13–14 tok/s without speculation — memory-bandwidth-bound for a 75GB model on the GB10. With DFlash speculation and coding prompts, testers report 22–24 tok/s. On an Apple M5 Max with 128GB unified memory, Ollama's Q4_K_M variant hits 49 tok/s at short context.

What We Don't Know Yet

• Independent third-party benchmark verification. All published scores come from Poolside's own pool harness. The company released full trajectories to address this, but community harnesses have shown variance — one Reddit tester noted 0.40 → 0.79 swings on schema tasks depending on max_tokens.
• Reliability of Terminal-Bench numbers under different harnesses. DeepSWE showed the widest gaps between vendor and community runs.
• Fabrication behavior under pressure. One private evaluation found the model scored 0.80 on grounding-under-pressure versus 0.97 for Qwen 3.5 122B, inventing figures and names in specific traps.
• RTX 6000 PRO stability. Poolside CEO Eiso Kant acknowledged bugs on that hardware in the first 24 hours, with fixes in progress.
• Intermediate thinking-effort control. Currently binary (on/off); Poolside lists finer control as a roadmap item.

Frequently Asked Questions

What is Laguna S 2.1?

Laguna S 2.1 is an open-weight Mixture-of-Experts model released by Poolside on July 21, 2026. It has 118 billion total parameters with 8 billion activated per token, supports a 1M-token context window, and is designed for agentic coding and long-horizon software engineering tasks.

Who made Laguna S 2.1?

Laguna S 2.1 was built by Poolside, a San Francisco AI lab led by co-CEOs Eiso Kant and Jason Warner. Poolside focuses on coding models for governments, defense agencies, and regulated enterprises, and has now released three models in three months.

Is Laguna S 2.1 open source?

Laguna S 2.1 is open-weight under the OpenMDW-1.1 license, which permits commercial use. Weights are published on Hugging Face at poolside/Laguna-S-2.1 in BF16, FP8, INT4, NVFP4, GGUF, and MLX formats, with a paired DFlash draft model for speculative decoding.

How much does Laguna S 2.1 cost?

Laguna S 2.1 costs $0.10 per million input tokens and $0.20 per million output tokens on OpenRouter's paid endpoint, with a cache-read rate of $0.01 per million. A free 262K-context tier is also available on OpenRouter and OpenCode; usage on the free tier may be used for training.

What hardware do I need to run Laguna S 2.1?

Laguna S 2.1 needs roughly 75GB of memory at Q4_K_M quantization, so a 128GB unified-memory machine such as an Apple Silicon Mac or NVIDIA DGX Spark is the practical minimum. It also runs on datacenter GPUs like the NVIDIA H200 and RTX 6000 PRO, where higher memory bandwidth yields much faster decode.

How does Laguna S 2.1 compare to DeepSeek V4 and Kimi K3?

Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, ahead of DeepSeek-V4-Pro-Max (64.0%) and Inkling (63.8%) but behind Kimi K3 (88.3%). It is dramatically smaller than all three — 118B total parameters versus 1.6T for DeepSeek-V4-Pro-Max and 2.8T for Kimi K3.

What benchmarks does Laguna S 2.1 lead?

Laguna S 2.1 posts 70.2% on Terminal-Bench 2.1, 78.5% on SWE-Bench Multilingual, 59.4% on SWE-Bench Pro (public), 40.4% on DeepSWE, 46.2% on SWE Atlas, and 49.7% on Toolathlon Verified. Poolside publishes full evaluation trajectories at trajectories.poolside.ai for third-party audit.

What to Watch Next

Three signals will decide how Laguna S 2.1 ages: independent Terminal-Bench and SWE-Bench runs from third-party leaderboards like Artificial Analysis, the resolution of the RTX 6000 PRO bugs that Poolside acknowledged on launch day, and whether the free OpenRouter tier survives real usage volumes without heavy rate limits. A finer thinking-effort control and vision support would also broaden the model's daily-driver appeal beyond pure coding.

Building similar agentic coding workflows? On kie.ai you can try Kimi K3, Claude Sonnet 5, and GPT-5.6.