DeepSeek V4 / V4.1 / MoE
Same footprint, new weights: a clean post-training comparison
New to model sizes or GPU memory? Start with weights, parameters and quantization →
Treat 1017 as a behavior update to V4.1 Flash, not a new deployment problem. Its configuration file is byte-identical and its tensor payload matches, so an existing V4.1 Flash plan carries over; the evaluation does not.
552B backbone only; 196B Engram plus vision and speculative components are separate.
Evaluate it for
Teams already serving V4.1 Flash who want better agent behavior without reworking memory placement.
Choose another path when
A plan that assumes a chat template file exists: this repository ships a Python encoding reference instead.
Release context
DeepSeek calls it a minor update with gains on reasoning, coding, agent and visual tasks, plus shorter outputs and task times. Because the shapes are unchanged, it isolates what post-training alone changes in quality and output length.
DeepSeek's update log and the weights repository share the date.
Release evidence · Announced 2026-10-17 · Weights 2026-10-17 · Follow the DeepSeek V4 / V4.1 timeline →
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.
Runtime recipe
GPUs within an NVLink system; this is not a count of whole racks.
vLLM 0.31.0+ (recipe default)
TP4 on one GB200 NVL4 tray. The recipe's verified numbers were taken on the September checkpoint, which shares this configuration and footprint; 1017 was not separately re-run.
deepseek-ai/DeepSeek-V4.1-Flash-1017
Checked 2026-11-06. Read the documented setup →
Runtime recipe
288 GB per GPU
vLLM ROCm nightly
TP2 with Engram tables offloaded to host memory; TP4 keeps Engram resident. Needs a ROCm build with the fix the recipe names; MI355X is not in vLLM 0.31.0.
deepseek-ai/DeepSeek-V4.1-Flash-1017
Checked 2026-11-06. Read the documented setup →
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.
A uniform-precision GPU count would hide this model’s component layout. 552B backbone only; 196B Engram plus vision and speculative components are separate. Use the documented artifact and placement path above; no GPU count is inferred here.
Architecture in practice
Publisher active path: 8B prefill / 16B decode.
552B backbone + 196B Engram is not one flat GPU budget.
Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.
| Exact checkpoint | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
|---|---|
| Text model type | deepseek_v41_text |
| Layers | 40 |
| Hidden width | 5,120 |
| Attention / KV heads | 64 / 1 |
| Head dimension | 512 |
| Routed / selected experts | 384 / 6 |
| Configured positions | 1,048,576 |
| Documented extension | Not recorded for this checkpoint |
| Inputs → output | text, image → text |
| License metadata | mit |
Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.
A useful comparison
The September V4.1 Flash is the controlled baseline; MiniMax M3.1 Flash is a different sparse-attention design aimed at similar agent work.
10 Sept 2026
The September model that changes the deployment diagram
552B reported total · text + image
23 Oct 2026
M3 with sparse attention everywhere and a quantized-trained store
428B 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 why identical configuration and payload let you reuse a capacity plan, and why shorter outputs can still change throughput and cost per task.
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, “DeepSeek-V4.1-Flash-1017: release, architecture and deployment”, checked 2026-11-06. Preserve this date and the exact checkpoint when citing.
