MoE enthusiast35B total, 3B active per token (MoE: 256 experts, 8 routed + 1 shared), hybrid Gated DeltaNet + Gated Attention layers. 256K native context, extensible to ~1M. The agentic flagship - ranked #1 local model for agent workflows in 2026 (perfect tool-calling, zero catastrophic failures on an 84-scenario / 16-category / 8-trial benchmark). Q4_K_M (~20GB) for GGUF runtimes; Q8_0 (~37GB) for full precision.
- 35.0B
- 262k
- apache 2.0
- Apr 2026
Scores #
Run it locally #
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
Can you run it? - reference rigs
| Rig | Q4_K_M | Q8_0 |
|---|---|---|
| NVIDIA Jetson Orin NX 16GB | ||
no -> cloudFit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
How can a 24GB GPU run a 744B model? It does not load the model into VRAM. The quantized weights (e.g. ~410GB at Q4) sit in system RAM; the GPU only holds the small shared attention and router tensors and accelerates prompt processing. Because GLM 5.2 is a Mixture-of-Experts model, each token activates only ~40B of its 744B params, so llama.cpp streams just those active experts from system RAM to the GPU each token (the -cmoe
offload path).
That makes decode speed bound by system-RAM bandwidth, not GPU bandwidth - single digits on DDR4, which is why these rigs show 3-8 t/s even though they “fit.” A bigger GPU (e.g. 2x 3090) keeps more experts resident on-card and raises tok/s; a smaller GPU still runs it but pays the bandwidth tax. A 744B dense model could not run this way - only MoE’s small-active-params trick makes it possible.
Aggressive quants (1-2 bit) trade accuracy for size - roughly 17% accuracy loss at 2-bit vs full precision, and real long-context work often needs Q5 or Q6 even when lower quants “fit.”
Formula estimates here are conservative; real tuned setups can exceed them (one HN user reports ~6 tok/s on a 512GB DDR4 + 2x 3090 rig).
Download options #
Inference cost over time #
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.