Laguna S 2.1 Free Is Now on OpenCode Zen — Why This Open-Weight Model Is a Big Deal

July 21, 2026

aimodelsopencodecodingopen-sourceagentsreview

The Release

Poolside released Laguna S 2.1 today – and OpenCode added it to their Zen free endpoint within hours. OpenCode announced it with: “Laguna S 2.1 is now free on OpenCode, 1M Context, fully open source. Poolside’s most capable model to date.”

This is the kind of release that doesn’t make front-page news but actually matters for developers who build with AI agents.

The Specs

Laguna S 2.1
Architecture MoE (Mixture of Experts)
Total params 118B
Active params 8B per token
Context window 1M tokens (1,048,576)
License OpenMDW-1.1 (open-weight)
Released July 21, 2026
Available on OpenCode Zen (free), OpenRouter (free), Hugging Face, Ollama
Local hardware Runs on single DGX Spark (NVFP4 quantized)

The headline stat: 8B active parameters. That’s tiny. For comparison, DeepSeek V4 Pro Max has 49B active. Claude Fable 5 has undisclosed but estimated 100B+. Nemotron 3 Ultra has 55B active.

Laguna S 2.1 does what it does with 8B active params. That’s the story.

Benchmarks

Here’s how it stacks up against models many times its size:

Terminal-Bench 2.1 (real-world terminal tasks)

Model Active Params Score
Kimi K3 50B 88.3
Claude Fable 5 ~100B+ 88.0
Laguna S 2.1 8B 70.2
Tencent Hy3 21B 71.7
Inkling 41B 63.8
DeepSeek V4 Pro Max 49B 64.0
Nemotron 3 Ultra 55B 56.4

Other Benchmarks

Benchmark Laguna S 2.1 DeepSeek V4 Pro Max Comparison
SWE-Bench Multilingual 78.5 76.2 +2.3 points
SWE-Bench Pro (Public) 59.4 55.4 +4 points
DeepSWE 40.4 9.0 +31.4 points
SWE Atlas (Codebase QnA) 46.2 27.2 +19 points
Toolathlon Verified 49.7 55.9 -6.2 points

The DeepSWE score is the standout. Laguna S 2.1 crushes DeepSeek V4 Pro Max (40.4% vs 9.0%) despite having 6x fewer active parameters. DeepSWE tests real-world, long-horizon software engineering tasks – the kind of work coding agents actually do.

On SWE-Bench Multilingual, it even edges out Qwen 3.7 Max (78.3) and DeepSeek V4 Pro Max (76.2).

Why the Active Parameter Count Matters

MoE models only activate a fraction of their total parameters per token. Laguna S 2.1’s 8B active means:

  1. Fast inference – fewer params to compute per token
  2. Run locally – fits on consumer hardware with quantization
  3. Lower cost at scale – 8B active = 1/6th the compute of a 49B active model
  4. Free endpoint is sustainable – Poolside can offer free access because it’s cheap to serve

Poolside specifically mentions it runs on a single NVIDIA DGX Spark with the official NVFP4 quantized version. That’s a $3,000 desktop machine, not a data center.

Open Weights + Free API = Unusual

Most model releases give you either open weights or a free API, rarely both.

You can:

  1. Use it for free today through OpenCode or OpenRouter – no credit card needed
  2. Run it locally on a DGX Spark or similar through Ollama
  3. Fine-tune it since the weights are open
  4. Build products on it under the OpenMDW-1.1 license

This combination is rare for a model that scores 40%+ on DeepSWE.

How It Compares to Other Free Models

On OpenCode Zen, the free tier now includes:

Model Context Best For
Big Pickle N/A General reasoning
Nemotron 3 Ultra 1M Heavy coding
DeepSeek V4 Flash 1M Speed
MiMo V2.5 1M Multimodal
North Mini Code 256K Light tasks
Laguna S 2.1 1M Agentic coding
Hy3 preview 256K Coding

Laguna S 2.1 fills the “agentic coding” slot that was previously missing from the free tier. The 1M context on a free endpoint is also unusually generous.

The Timing

This launched on the same day as Gemini 3.6 Flash. Two coding model releases on the same day, but very different approaches:

For local/offline agent setups, Laguna S 2.1 is the more interesting option because you can actually run it yourself.

The Bottom Line

I switched my coding agent’s secondary model to Laguna S 2.1 on OpenCode Zen free to test it. For a model with 8B active parameters, the coding quality is surprising. The DeepSWE score against larger models suggests Poolside optimized specifically for the kind of work coding agents do – long tool call chains, debugging, multi-file edits.

If you run coding agents and haven’t tried it, it’s worth a test drive. Free endpoint, 1M context, no credit card. Hard to beat that price.

FAQs

Q: Is Laguna S 2.1 really free?

A: Yes. OpenCode Zen offers it free (no credit card needed). OpenRouter also has a free endpoint with 262K context. Both are legitimate free tiers, not trials.

Q: Can I run Laguna S 2.1 locally?

A: Yes. It’s on Ollama and Hugging Face. With 8B active params, it can run on a single high-memory machine. The official NVFP4 quantized version runs on a DGX Spark.

Q: How does it compare to DeepSeek V4 Flash?

A: Different strengths. DeepSeek V4 Flash is faster. Laguna S 2.1 is better at complex multi-step coding tasks and has a much higher DeepSWE score.

Q: What’s the license?

A: OpenMDW-1.1 from Poolside. Open-weight, allows commercial use. Check the license terms on Hugging Face for specifics.

Q: Is it better than Gemini 3.6 Flash?

A: On raw coding benchmarks, Gemini 3.6 Flash scores higher (49% vs 40.4% on DeepSWE). But Gemini is proprietary and costs $7.50/1M output. Laguna S 2.1 is open-weight and free. Different tools for different constraints.

Q: 8B active parameters seems small. Does it actually work well for complex tasks?

A: Yes – that’s the surprising part. The benchmarks show it beating models with 5-6x more active parameters. The MoE architecture lets it be small at inference time while retaining high capability.


Related: Gemini 3.6 Flash Released – also launched today. AI Agents for Developers – my current agent stack and how I choose models. OpenCode vs Claude Code – comparing the two coding agent platforms.