24 Feb 2026
Qwen3.5-35B-A3B
The easiest MoE memory lesson to make concrete
35B reported total · text + image + video
GLM / MoE
GLM-4.7-Flash
A compact way to learn latent-attention serving
New to model sizes or GPU memory? Start with weights, parameters and quantization →
This 30B/3B MoE is a useful GLM entry point when the later 744B-class models are far beyond your hardware. Its latent attention makes it a better systems lesson than a generic small-model comparison.
30Breported total parameters
3Bactive parameters per token
202,752configured token positions
Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Evaluate it for
Smaller text coding experiments and study of latent-attention caches.
Choose another path when
Assuming an upgrade to GLM-5.3-Flash keeps the same hardware footprint.
Release context
Why this release matters
The January release occupies a different deployment class from the later 320B GLM-5.3-Flash. “Flash” is a product-family label, not a stable size or latency category.
This is the dated Z.ai release-note event; a separate first weight-upload day has not been certified.
Deployment starting points
Which hardware, how many, at what precision?
Start with an exact artifact and an engine that supports it. A configuration below is evidence of a documented path; its device count does not promise a particular throughput or concurrency.
No certified hardware configuration in this edition
Assuming an upgrade to GLM-5.3-Flash keeps the same hardware footprint.
The pinned model card establishes the architecture. It does not, by itself, prove that a particular GPU count runs this checkpoint at your target context. The capacity screen below helps rule out undersized allocations; a successful load and workload test are still required.
31.22 GB of tensor data in the pinned index
This is the index’s reported tensor payload, in decimal GB. It excludes file headers and runtime memory. Mixed precision, conversion and host/device placement determine how much GPU memory the loaded model needs.
Inspect the exact index metadata →Capacity screen: would the weights fit?
This estimates inference memory for a hypothetical uniform precision. It is useful for rejecting an allocation that is too small. It does not establish a working deployment or estimate training memory.
| Representation | Approximate weight floor | What remains to budget |
|---|---|---|
| 16-bit uniform | 60 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 8-bit uniform | 30 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 4-bit uniform | 15 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
Enter the total across the whole model replica, not per GPU: KV or recurrent cache for all concurrent requests, image/audio state and execution workspace. There is no universal default. We additionally leave 15% of each GPU unused.
60 GB for hypothetical weights
Enter a request-and-buffer budget to calculate the aggregate capacity requirement.
A count of one does not prove the engine supports this GPU. A count above one does not prove the model shards evenly or communicates efficiently. Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Architecture in practice
What the serving engine has to do
Most experts are stored. Only a few execute.
Compute follows selected experts.
Weight memory follows the full expert store.
Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.
Inspect the full checkpoint specifications
| Exact checkpoint | zai-org/GLM-4.7-Flash |
|---|---|
| Text model type | glm4_moe_lite |
| Layers | 47 |
| Hidden width | 2,048 |
| Attention / KV heads | 20 / 20 |
| Head dimension | Not verified |
| Routed / selected experts | 64 / 4 |
| Configured positions | 202,752 |
| Documented extension | Not recorded for this checkpoint |
| Inputs → output | text → text |
| License metadata | mit |
Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.
A useful comparison
What else belongs on the shortlist?
Qwen provides multimodal hybrid attention; gpt-oss provides a compact mixed-precision reasoning alternative.
5 Aug 2025
gpt-oss-20b
A compact reasoning baseline that still earns its place
21B reported total · text
These are editorial comparison candidates. We have not run a matched quality or serving benchmark, so this is not a ranking.
Prepare to explain it
The interview lesson
Explain how latent cache storage differs from ordinary K/V tensors, and why the active count still does not determine weight memory.
Attention layouts · Quantization · Courses and worked examples
For a deployment evaluation, record exact weights, precision, engine, device count, interconnect, prompt/output lengths and concurrency. Report task success, errors, TTFT, TPOT and useful throughput together.
Sources and citation
Our assessment is editorial judgment based on the linked architecture and deployment evidence. Model facts are publisher-reported or attributed to runtime maintainers; no independent GPU benchmark was run.
AI Infra Interviews, “GLM-4.7-Flash: release, architecture and deployment”, checked 2026-09-13. Preserve this date and the exact checkpoint when citing.