$ cat wiki/models/laguna-s-2-1.md
Laguna S 2.1
Compared with
Spec
| Attribute | Value |
|---|---|
| Developer | Poolside |
| Released | 2026-07-21 |
| Announced | 2026-07-21 |
| Context window | 1M |
| Pricing | unknown |
| License | OpenMDW-1.1 (open-weight) |
| Availability | Hugging Face |
Pricing is unknown: Poolside published weights, not an endpoint price, and the | |
| model appears on neither leaderboard snapshot this repo holds | |
| (source). |
| Attribute | Value |
|---|---|
| Catalogue id | unknown |
Release Date
2026-07-21 (source).
118B total parameters, 8B active per token — a Mixture-of-Experts built for agentic coding, with a context window up to 1M tokens in both thinking and no-thinking modes (source).
The license is OpenMDW-1.1, which permits use, modification and redistribution including commercial use (source).
Benchmarks
These are Poolside's own figures on Poolside's own compiled leaderboard, and no harness configuration is published alongside them. Per Eval Harness Configuration they are claims about a (model, harness) pair rather than properties of the model (source):
| Benchmark | Laguna S 2.1 | Poolside's stated comparison |
|---|---|---|
| Terminal-Bench 2.1 (thinking) | 70.2% | first among open, disclosed-size models on their leaderboard; behind only larger or closed systems |
| SWE-Bench Multilingual | 78.5% | tops their published table outright |
| DeepSWE v1.1 | 40.4% | against DeepSeek-V4-Pro-Max at 9.0%, at roughly one-sixth the active parameters |
| Coverage reports that on long-horizon coding benchmarks the model holds its own | ||
| against systems several times its size, naming DeepSeek-V4-Pro-Max, NVIDIA | ||
| Nemotron 3 Ultra and Inkling | ||
| (source). |
No independent measurement of any of this exists in the wiki. Laguna S 2.1 is absent from all 260 rows of the Artificial Analysis leaderboard read 2026-08-02 (source).
Use Cases
Agentic coding — long-horizon software tasks, which is what the benchmark set Poolside chose (Terminal-Bench, SWE-Bench Multilingual, DeepSWE) measures (source). See Agentic Reinforcement Learning and Agents (LLM Agents).
Compared To
- DeepSeek V4-Flash — 284B/13B, MIT, also re-post-trained for agentic coding; the direct point of comparison for "open coding model at small active size"
- Inkling — 975B/41B, Apache 2.0; the generalist to Laguna's specialist, and one of the models Poolside's coverage benchmarks against (source)
- Kimi K3 — 2.8T; the size end of the same open-weight window
- Mistral Large 3 — the previous Western frontier-scale open-weight release this wiki tracks
Conflicting Reports
Headlines disagree on the size multiple Laguna S 2.1 is said to beat. VentureBeat's headline says "rivals 10x its size"; another outlet's says "14x its size" (source). Both are paraphrases of the same vendor comparison table rather than separate measurements, and neither multiple appears in the figures Poolside published. The wiki states the figures instead.