Two new architectures landed in our engine this week. Both are over 100B parameters.

July 30, 2026 · 5 min read

llama.cpp shipped two new architecture tags in the last two days — GLM_DSA for Zhipu's GLM-5.2 (tag b10174, July 29) and one for Poolside's Laguna-S-2.1 (tag b10173, July 28). Both are real, both are open-weight, and both are useless to a phone: GLM-5.2 is roughly 754B total parameters with about 40B active per token (MoE), and Laguna-S-2.1 is 118B total with 8B active. Nothing about "active parameters" shrinks the download — the full expert set still has to live somewhere. Zero catalog changes today, for the same size reason as most weeks.

What actually shipped

ModelVendorParams (total / active)LicenseEngine support
GLM-5.2Zhipu (Z.ai)~754B / ~40BOpen weightllama.cpp b10174, Jul 29
Laguna-S-2.1Poolside118B / 8BOpenMDW-1.1llama.cpp b10173, Jul 28

Sources: ggml-org/llama.cpp releases; poolside.ai/blog/introducing-laguna-s-2-1.

Laguna-S-2.1 is the more interesting of the two on paper — Poolside ships official GGUF and MLX builds and pitches it as runnable on a single NVIDIA DGX Spark, which is a real local-deployment story. It's also 118B parameters, which is a DGX-Spark-and-256GB-of-unified-memory story, not a phone story. Both models land squarely in what the engine can now load, not what our 2.5 GB, phone-download catalog can use. That gap — "the engine caught up" and "the catalog grew" being two completely different sentences — has been true for months and stays true this week.

The one model worth actually reading about this week doesn't load at all

Nanbeige4.2-3B, published July 24, is a 3B-parameter model that beats Qwen3.5-9B and Gemma4-12B — three to four times its size — on GPQA Diamond, SWE-Bench Verified, and a math benchmark (HMMT-Feb-2026) by as much as 31 points. It does it by looping its own transformer layers twice per forward pass instead of adding parameters. It's Apache 2.0, a Q4_K_S build just clears our 2.5 GB budget, and it currently runs in exactly zero apps built on mainline llama.cpp, Ollama, or LM Studio — including this one — because that looping computation isn't a graph shape the mainline engine builds yet. We wrote the full story, with the benchmark table and the closed GitHub issue asking for support, here.

Catalog: unchanged

20 models, same 9 categories as SLM Daily #5. Nothing added, nothing removed today. G9v3-3B (added July 27) and that edition's catalog fix — removing the oversized Qwen3.5 4B and IBM Granite 3.1 8B entries — both stay as they were. If Nanbeige4.2-3B's architecture lands in a mainline llama.cpp release, it goes to the top of the list for next time.

Discuss this on the forum →