Laguna S 2.1
Poolside · Code · MoE 118B (A8B) · 1049k context · Released 24 August 2026
Laguna S 2.1 is a sparse mixture-of-experts built for coding: 256 routed experts (top-ten) plus one shared, across 48 layers with mixed sliding-window and global attention, offering thinking and non-thinking modes and a 1-million-token context. It is text-only. The notable thing is that it is genuinely open: the full BF16 weights, plus official NVFP4 and GGUF builds, are on Hugging Face rather than behind an API, under the permissive OpenMDW licence. At 118 billion total parameters it is server-scale: even a 4-bit build is around 70GB, so it wants a 128GB unified-memory machine or a multi-GPU server, with the usual mixture-of-experts caveat that the total, not the 8 billion active, sets the footprint.
Strengths
- Genuinely open weights under a permissive licence, from a commercial coding lab
- A dedicated coding model with a 1-million-token context and thinking modes
- Sparse: only about 8 billion of 118 billion parameters compute per token
- Official 4-bit (NVFP4) and GGUF builds published alongside the full weights
Weaknesses
- Server-scale: ~70GB at 4-bit and ~235GB at BF16, so not a consumer-card model
- Text-only and coding-focused, not a general assistant
- Community reports of occasional loops and stability issues
- Vendor benchmark claims await independent verification
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| NVFP4 / Q4 (as released) | ~71GB | Official NVFP4 build is roughly 71GB; needs a 128GB unified-memory machine or multi-GPU. Driven by the 118B total, not the active count. Approximate. |
| Q8 (approx) | ~120GB | Roughly 120GB. Multi-GPU or server territory. Approximate |
| BF16 (as released) | ~235GB | The full release, roughly 235GB across 46 shards. Server territory. Approximate |
What you'd need to run this
Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.
Minimum to run it
NVFP4 / Q4 (as released) · ~71GB needed
96GB of unified memory
AMD Ryzen AI Max+ 395 (Strix Halo)or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA RTX PRO 6000 Blackwell .
Unified memory is shared with the model, so it is already counted above.
For good quality
Q8 (approx) · ~120GB needed
A multi-GPU server, roughly 2× 80GB-class GPUs
Server memory in the hundreds of gigabytes or more, plus fast interconnect.
tens of thousands of pounds, or rented by the hour
Usually rented — see the cost calculatorBest quality
BF16 (as released) · ~235GB needed
A multi-GPU server, roughly 4× 80GB-class GPUs
Server memory in the hundreds of gigabytes or more, plus fast interconnect.
tens of thousands of pounds, or rented by the hour
Usually rented — see the cost calculatorLicence
OpenMDW-1.1 — read the licence
Laguna S 2.1: common questions
- What hardware do I need to run Laguna S 2.1?
- At its most compressed (NVFP4 / Q4 (as released)) it needs roughly 71GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is Laguna S 2.1 free for commercial use?
- Yes. Laguna S 2.1 is licensed under OpenMDW-1.1, which permits commercial use with no meaningful conditions.
- Can I run Laguna S 2.1 on Apple Silicon?
- It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
- What is Laguna S 2.1's context window?
- Laguna S 2.1 has a context window of 1,048,576 tokens, about 1049k.
Availability
Recommended for
- Server-class or 128GB-unified coding work kept in-house
- Long-context coding across large repositories
- Teams wanting a permissively licensed dedicated coding model
Related models
Related guides
Glossary
Catalogue entry last verified 2 October 2026. Specifications change; verify anything you are about to spend money on.