#3593 [Engine] Laguna 2.1 family support: S 2.1 (118B-A8B) and XS 2.1 (33B-A3B)

本文

### What do you want to do?

Run the Laguna 2.1 family on Strata: **Laguna S 2.1** (118B-A8B, 1M context) and **Laguna XS 2.1** (33B-A3B) from poolside.ai (2026-07-21).

Who needs it: owners of 128 GB desktop DDR4 + a single 16-24 GB consumer GPU who want an open-weight agentic-coding model with native 1M context. The architecture looks tailor-made for the RAM-to-VRAM expert hierarchy Strata already ships:

- **MoE layout matches Strata's sweet spot**: 47 MoE layers x 256 routed experts (top-10, token-choice softplus router) + 1 shared expert, `moe_intermediate_size=1024` — small per-expert footprints (~9.4M params each, ~12k routed experts), the exact regime where the existing qwen-family hierarchy excels.
- **Hybrid attention makes long context cheap**: 12 full GQA layers (8 KV heads, head_dim 128) + 36 sliding-window layers (window 512). Global KV is ~24 KB/token, so a 1M-context session is mostly RAM-resident KV with a small VRAM window — cheaper KV than any family Strata ships today.
- **Simpler than the qwen family**: no PLE / n-gram tables, so the plain-MoE path applies. **No MTP head**, but Poolside ships a *trained DFlash drafter* (`poolside/Laguna-S-2.1-DFlash`, 2.2 GB) for speculative decoding instead.
- **Weights and GGUFs are fully open, no gating**: license OpenMDW-1.1; `laguna` arch in llama.cpp (upstream b10087+, `poolsideai/llama.cpp@laguna` branch with DFlash spec); Unsloth Dynamic 2.0 quants (UD-IQ4_XS 57.6 GB, UD-Q4_K_M 73.1 GB, Q5/Q6/Q8), official poolside GGUFs, APEX quants for XS.

Happy to provide logs and test data from a 5070 Ti + 128 GB DDR4 machine and to run benches against candidate builds.

### What do you do today instead?

llama.cpp on the `laguna` arch (upstream b10087+ / poolside fork) with `-cmoe`/`--n-cpu-moe`, mmap of the GGUF on NVMe and the DFlash drafter for spec decode. On a 5070 Ti 16 GB + 128 GB DDR4 box UD-Q4_K_M (73.1 GB) loads and runs, but decode is bound by plain RAM streaming of ~2.2 GB/token of expert weights in Q4, with no hierarchical expert cache and no VRAM expert tier — a textbook 2-3x Strata-style win. We are running exactly such a llama.cpp llama-server bench on that machine right now (UD-Q4_K_M) and will add the measured decode/prefill numbers as a comment here once it completes.