24 Feb 2026
Qwen3.5-35B-A3B
The easiest MoE memory lesson to make concrete
35B reported total · text + image + video
Qwen / MoE
A middle step before the largest Qwen3.5 model

New to model sizes or GPU memory? Start with weights, parameters and quantization →
The 122B variant lets you test whether a larger expert pool improves your difficult tasks without moving immediately to 397B. Demand a measurable quality gain over 35B-A3B before accepting its storage and sharding costs.
Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Evaluate it for
Multi-GPU multimodal evaluation where the 35B model misses important cases.
Choose another path when
An application whose error rate is already acceptable on the smaller model.
Release context
The February expansion filled the gap between the first 397B release and small-device models. Ten billion active parameters describes the selected path, not the resident model.
Deployment starting points
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.
An application whose error rate is already acceptable on the smaller model.
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.
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 →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 | 244 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 8-bit uniform | 122 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 4-bit uniform | 61 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.
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
Hybrid attention does not mean zero request state.
A full-attention formula cannot describe every layer.
Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.
| Exact checkpoint | Qwen/Qwen3.5-122B-A10B |
|---|---|
| Text model type | qwen3_5_moe_text |
| Layers | 48 |
| Hidden width | 3,072 |
| Attention / KV heads | 32 / 2 |
| Head dimension | 256 |
| Routed / selected experts | 256 / 8 |
| Configured positions | 262,144 |
| Documented extension | 1,010,000 tokens; additional configuration and evaluation required |
| Inputs → output | text, image, video → text |
| License metadata | apache-2.0 |
36 linear attention layers; 12 full attention layers. These counts describe the configured pattern.
Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.
A useful comparison
These bracket its storage class and make the scale-up decision testable.
24 Feb 2026
The easiest MoE memory lesson to make concrete
35B reported total · text + image + video
16 Feb 2026
The launch model that established Qwen3.5’s hybrid design
397B reported total · text + image + video
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
Explain how expert placement and cross-device communication change when a model stops fitting on one device.
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.
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, “Qwen3.5-122B-A10B: release, architecture and deployment”, checked 2026-09-13. Preserve this date and the exact checkpoint when citing.