Z.ai's GLM family runs from the 110B GLM-4.5-Air and 31B GLM-4.7-Flash MoEs up to the 754B GLM-5.1/5.2 flagships with up to 1M context. All four catalog models are MIT licensed.
Z.ai (智谱 / Tsinghua spin-out) is the force behind the GLM family, and every model on this page carries the same MIT license — a consistent permissive stance that sets the family apart from competitors with custom terms. The lineup has four entries organized across two generations: GLM-4 (the 4.5 Air and 4.7 Flash) and GLM-5 (5.1 and 5.2). All four are MoE models, so active parameters stay far smaller than the headline numbers, and context windows reach up to 1M tokens on the top models.
VRAM estimates below use Q4 weights + KV cache at 8K context + 1 GB overhead. For MoE models the KV cache tracks active parameters, not total — which makes GLM-4.7-Flash especially efficient at long context despite storing 31.2B parameters.
Size ladder
- GLM-4.7-Flash (31.2B MoE, 3B active, 198K ctx) — Q4 ≈ 17.2 GB → 24GB card (RTX 3090/4090 class); only 3B active per token makes it the fastest GLM in our catalog, in the same ballpark as the popular A3B designs from the Qwen family.
- GLM-4.5-Air (110.5B MoE, 12B active, 128K ctx) — Q4 ≈ 60.8 GB → just beyond 48GB; needs a 96GB-class card at Q4 or Q2 on 48GB (≈31 GB weights).
- GLM-5.1 (753.9B MoE, 40B active, 198K ctx) — Q4 ≈ 414.7 GB → multi-GPU server, roughly 5×96GB; the GLM-5 generation generalist with near-200K context.
- GLM-5.2 (753.3B MoE, 40B active, 1M ctx) — Q4 ≈ 414.3 GB at 8K → same multi-GPU tier as 5.1, but the 1M context window means long-document KV cache adds up — roughly 100 GB to run the full window at Q4.
Family highlights
- Surgically small active params. GLM-4.7-Flash activates only 3B per token inside 31.2B total — that is a 10:1 ratio and one of the smallest active footprints we track. On a 24GB card it runs with near-200K context headroom because the KV cache barely grows.
- GLM-5: big models, MIT license. A 754B MoE under MIT terms is unusual. Most families with models this size use custom or restricted licenses. GLM-5.1 and 5.2 are genuinely free to use commercially, fine-tune, and redistribute — the limiting factor is hardware, not legal terms.
- Context jumps from 128K to 1M across generations. GLM-4.5 sits at 128K, 4.7 bumps to 198K, and GLM-5.2 hits 1M. GLM-5.1's 198K is generous for RAG and agent workloads without the extreme cache cost of 1M.
- GLM-4.5-Air is the bridge. At 110.5B total / 12B active, it is the mid-size option between the portable 4.7-Flash and the server-class 5.x flagships. Q4 on a 96GB card is the natural home.
Which one should you pick
- 12GB / 16GB VRAM — neither 4.7-Flash (Q4 ≈ 17.2 GB) nor 4.5-Air fit. GLM is a 24GB-and-up family.
- 24GB VRAM — GLM-4.7-Flash at Q4 is the only option, and it is a good one: 3B-active speed, up to 198K context, MIT licensed. This is the practical GLM pick for a single consumer GPU.
- 48GB VRAM — 4.7-Flash runs with room to spare. 4.5-Air becomes possible at Q2 (≈31 GB weights); Q4 still does not fit.
- 96GB VRAM — GLM-4.5-Air at Q4 (≈60.8 GB) with comfortable headroom for 128K context. 4.7-Flash is trivial here.
- Multi-GPU (5×96GB+) — GLM-5.1 and 5.2 at Q4. For 5.2 specifically, budget 100+ GB extra for the full 1M context window.
Most single-GPU users will land on GLM-4.7-Flash on a 24GB card: it gives you the family's MoE speed, MIT licensing, and a generous 198K context at under 18 GB. GLM-4.5-Air is the upgrade for 96GB rigs; GLM-5 is firmly datacenter territory.
Links
4 models in the GLM family, grouped by series. Q4 (GB) is weights-only; total VRAM adds KV cache and overhead — the fit check links compute it for a representative retail GPU at 8K context.