GLM-4.7-Flash

What this model is

Z.ai describes GLM-4.7-Flash as a 30B-A3B mixture-of-experts model. The 47-layer checkpoint selects four of 64 routed experts plus one shared expert per token.

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-4.7-Flash is the consumer-friendly member of the family: 30B stored parameters, 3B active per token, and local Ollama Q4/Q8 tags. The catch is capacity rather than speed. Q4_K_M plus an 8K MLA cache and overhead is about 20.3 GB, so a 24 GB GPU is tight unless it drops to Q3_K_M. A 32 GB card or a 24 GB Mac using the lower-bit build is a more realistic local setup.

Z.ai documents vLLM, SGLang, and Transformers. Ollama, llama.cpp, and community MLX builds provide practical local routes, but verify the selected GGUF or MLX package rather than assuming every runtime uses the same kernels. The community MLX 4-bit package is about 16.85 GB and fits a 24 GB Mac with more headroom than the GGUF Q4 file.

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 GGUF repository. Ollama publishes matching local Q4_K_M and Q8_0 tags; the MLX 4-bit conversion is a separate 16.85 GB package.

Estimate formula: published weight files + 0.4435 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 16.6 GB before OS and application reserve.

Conservative FP16 MLA estimate from the 47-layer config: 512 compressed KV dimensions plus 64 RoPE dimensions per token. It excludes the MTP layer and runtime workspace.

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-4.7-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 measurement is claimed. The 3B active path should be fast for its class, but expert routing, 4-bit kernel support, and available memory bandwidth determine actual decode speed.

Active path: 3B 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

• Ollama
• llama.cpp
• MLX
• vLLM
• SGLang

ollama run glm-4.7-flash:q4_K_M

lms get "GLM-4.7-Flash"@q3_k_m
lms load <model-key-from-lms-ls> --gpu max --context-length 8192

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 30B stored experts determine memory, not the 3B active path. At 8K, Q4_K_M plus KV cache and overhead is about 20.3 GB, which is tight on a 24 GB GPU and does not fit a 16 GB device.
• Use Q3_K_M when 24 GB is a hard ceiling. The lower-bit file trades quality for enough headroom to avoid memory pressure.
• Official runtime documentation covers vLLM, SGLang, and Transformers. Ollama, llama.cpp, and community MLX builds are useful local paths but are not the only upstream-supported runtimes.

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-4.7-Flash
• https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF
• https://ollama.com/library/glm-4.7-flash
• https://huggingface.co/mlx-community/GLM-4.7-Flash-4bit
• https://huggingface.co/lmstudio-community/GLM-4.7-Flash-GGUF
• https://huggingface.co/zai-org/GLM-4.7-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