27 Jul 2026
Kimi-K3
A cluster model with a new attention design
2,800B reported total · text + image + video
Qwen / MoE
Qwen3.8-2.4T-A95B
Large-scale text reasoning, with large-scale residency
New to model sizes or GPU memory? Start with weights, parameters and quantization →
Treat this as a text-only cluster candidate. The useful comparison with K3 is about your coding and reasoning tasks, license constraints and operating complexity, not which headline parameter count is larger.
2,400Breported total parameters
95Bactive parameters per token
262,144configured token positions
Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Evaluate it for
Cluster-scale text reasoning and coding evaluation with a carefully pinned checkpoint.
Choose another path when
Image or video inputs, consumer hardware, or a project that cannot accept the checkpoint’s custom terms.
Release context
Why this release matters
The open weights arrived after the hosted Max announcement. The downloadable checkpoint is text-only; hosted Max features should not be copied onto its card.
The hosted Max announcement on August 2 does not date this downloadable checkpoint.
Release evidence · Weights 2026-08-12 · Follow the Qwen timeline →
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
Image or video inputs, consumer hardware, or a project that cannot accept the checkpoint’s custom terms.
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.
4,892.37 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 | 4,800 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 8-bit uniform | 2,400 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 4-bit uniform | 1,200 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.
4,800 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
Two kinds of layers retain two kinds of state.
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.
Inspect the full checkpoint specifications
| Exact checkpoint | Qwen/Qwen3.8-2.4T-A95B |
|---|---|
| Text model type | qwen3_5_moe_text |
| Layers | 92 |
| Hidden width | 8,192 |
| Attention / KV heads | 64 / 4 |
| Head dimension | 256 |
| Routed / selected experts | 512 / 10 |
| Configured positions | 262,144 |
| Documented extension | 1,010,000 tokens; additional configuration and evaluation required |
| Inputs → output | text → text |
| License metadata | qwen3.8-max · custom terms |
69 linear attention layers; 23 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
What else belongs on the shortlist?
K3 provides a multimodal cluster comparison; Qwen3.8-27B tests whether the large deployment is necessary.
14 Aug 2026
Qwen3.8-27B
A current dense baseline before you scale out
27B 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
The interview lesson
Explain model identity versus service identity, then account for all 2.4T resident parameters before discussing the 95B active path.
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, “Qwen3.8-2.4T-A95B: release, architecture and deployment”, checked 2026-09-13. Preserve this date and the exact checkpoint when citing.