GLM-5.3-Flash

What this model is

Z.ai reports 320B total parameters and 18B active parameters. The 45-layer hybrid model combines 34 linear-attention layers with 11 sparse-attention layers, and it is natively multimodal rather than a text-only Flash variant.

The architecture is mixture of experts. Memory must hold the full weight set, while active parameters affect the speed heuristic. Experts can be offloaded to system memory, but that trades speed for capacity.

How this GLM model actually runs locally

GLM-5.3-Flash is the most practical member of this new batch because its 18B active path is far smaller than the 320B stored expert set. It is still not a normal small-model download: the lowest published GGUF in this record is about 98.7 GB with the multimodal projector. A 128 GB Mac or unified-memory box can load the 1-bit-class build, but the result is a slow, low-bit specialist configuration. A 256 GB machine is the first generous local target in this catalog.

Z.ai's official card lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth. Llama.cpp has merged GLM-5.3-Flash architecture support in PRs 27773 and 31009. Ollama's current library entry is cloud-only; the site therefore does not generate a local ollama run command. A community 4-bit MLX package exists, but its model card says it requires mlx-vlm rather than mlx-lm and at least 256 GB of Mac memory. That is a specialist path, not evidence that every Apple Silicon Mac can run the model.

Runtime support was verified on 2026-11-02. Exact package names and kernel support can change quickly; use the source links below for the model, quantization, and runtime projects.

Memory and quantization

Decimal GB from Unsloth's published sharded GGUF files on 2026-11-02. Each total includes the 1.128 GB F16 multimodal projector. These files require a recent llama.cpp build, Unsloth Desktop, or a server runtime listed on the official card.

Estimate formula: published weight files + 0.8305 GB per 8K tokens × context / 8K + 1.5 GB runtime overhead. Sizes are decimal GB, not GiB. At 8K context, the smallest listed build is about 101.0 GB before OS and application reserve.

Conservative FP16 estimate from the 11 sparse-attention layers. It includes the 512-wide compressed KV latent plus the 32-head, 128-wide indexer cache; linear-attention state is excluded because it does not grow linearly by token.

Ollama tags were checked against the linked Ollama library pages on 2026-11-02. LM Studio's lms get suffix selects a file when the search results expose that exact quantization; confirm the local key with lms ls.

KV method source: https://huggingface.co/zai-org/GLM-5.3-Flash/raw/main/config.json

Speed heuristic

The model page does not have one universal speed number: memory capacity and bandwidth belong to the machine, not the checkpoint. Open a model-and-hardware page for the bandwidth-based range and its conditions.

No universal local measurement is claimed. The 18B active path is much smaller than the 320B stored set, but expert routing, CPU offload, and the sparse-attention kernels dominate real throughput.

Active path: 18B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark. No verified measurement is substituted for an estimate.

Public speed checks for this model

No verified measurement yet for this exact model record.

If the full weights do not fit in accelerator memory but do fit in system memory, the calculator marks the result as a tight CPU-offload fit. Expect substantially slower generation and verify the runtime's MoE offload support.

Runtimes

• llama.cpp
• vLLM
• SGLang

No local Ollama tag. Use a current llama.cpp build, vLLM, or SGLang; Ollama currently offers GLM-5.3-Flash only as a cloud model.

License restrictions

MIT permits commercial use, modification, and redistribution; preserve the copyright and permission notice.

This is a plain-language summary, not legal advice. Read the linked license before commercial deployment or redistribution.

Known issues

• The practical floor is about 100 GB of memory for the smallest 1-bit-class GGUF plus the projector, KV cache, and runtime workspace. A 128 GB machine can load it, but this is a low-bit, slow, and sensitive configuration rather than a normal 24 GB desktop workload.
• Ollama's glm-5.3-flash library entry is cloud-only. This site does not generate an ollama run command for it.
• A community 4-bit MLX conversion exists, but its card says mlx-lm does not support the glm5_next architecture and requires mlx-vlm. It also needs at least 256 GB of Mac memory, so it is not treated as a generic MLX runtime here.
• The 1M-token context is a capacity ceiling, not a practical default. Long contexts can still raise KV-cache and indexing memory beyond the 100 GB weight floor.

If it does not fit, try these smaller records

These alternatives have a lower recommended memory target in the current catalog. Open each page to compare the quantization-specific fit before downloading.

FAQ

Model-specific questions

Sources

• https://huggingface.co/zai-org/GLM-5.3-Flash
• https://z.ai/blog/glm-5.3-flash
• https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF
• https://github.com/ggml-org/llama.cpp/pull/27773
• https://github.com/ggml-org/llama.cpp/pull/31009
• https://huggingface.co/mlx-community/GLM-5.3-Flash-4bit
• https://ollama.com/library/glm-5.3-flash
• https://huggingface.co/zai-org/GLM-5.3-Flash/raw/main/config.json

This page is an independent compatibility guide, not official documentation. Verify the primary source before downloading or deploying a model.

Check a specific machine