The phrase “deepseek-v3 1 paper” is ambiguous: it could mean a paper about DeepSeek-V3.1, or one of the papers about DeepSeek-V3. The official material identified here includes a V3.1 release announcement, but this guide does not establish whether a standalone V3.1 technical paper exists. Don’t treat a paper about V3 as if it were about V3.1.
For the original DeepSeek-V3, two sources serve different purposes: read the DeepSeek-V3 Technical Report for the model’s architecture and training, and “Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures” for hardware and system co-design. Neither should be presented as a V3.1 paper.
Which DeepSeek paper should you read?
| If you’re looking for… | Read… | What it covers |
|---|---|---|
| DeepSeek-V3’s model design and training | DeepSeek-V3 Technical Report | The V3 technical report. |
| Hardware and infrastructure lessons from V3 | Insights into DeepSeek-V3 | Hardware and model co-design, including memory, compute, and interconnects. The authors distinguish it from the technical report. |
| DeepSeek-V3.1 release information | DeepSeek’s V3.1 release announcement | Release information—not a V3.1 research paper. |
The hardware paper is listed for ISCA ’25 and gives DOI 10.1145/3695053.3731412. It points readers to the separate technical report for detailed V3 architecture and algorithms.
What the V3 hardware paper says—and what it doesn’t
The ISCA paper focuses on hardware-aware design, not a broad comparison of how well V3 answers everyday prompts. It discusses memory-efficient attention (MLA), Mixture-of-Experts (MoE), FP8 training, inference systems, and networking. Those details can explain design trade-offs; they do not establish the quality, latency, or price you’ll get from a particular hosted API.
A few numbers illustrate why conditions matter:
- The paper reports 671 billion total parameters and 37 billion activated per token for V3. Its training-compute comparison gives about 250 GFLOPS per token, assuming a sequence length of 4,096. These are paper-reported model and compute figures, not API response-time guarantees.
- In a table labeled BF16 precision, it reports KV-cache use of 70.272 KB per token for DeepSeek-V3 with MLA, versus 327.680 KB for Qwen-2.5 72B with GQA and 516.096 KB for LLaMA-3.1 405B with GQA. Keep the precision, model, and attention-method labels attached to the comparison.
- The authors report that multi-token prediction (MTP) achieved 80%–90% acceptance for the second subsequent token and 1.8× generation TPS versus a setup without MTP. The paper’s report of these figures does not specify a complete test setup, so treat them as the authors’ reported results—not a general speed promise.
The paper also calculates a theoretical GB200 NVL72 throughput estimate above 1,200 tokens per second, but explicitly says it has not been empirically validated. Do not cite that estimate as observed production throughput.
What’s confirmed about V3.1?
DeepSeek’s official V3.1 release announcement describes hybrid Think and Non-Think modes and says the model received post-training for tool use and multi-step agent tasks. The announcement is useful for understanding the release, but it is not a substitute for a V3.1 research paper or a task-specific independent evaluation.
If you specifically need a V3.1 paper, don’t cite the original V3 technical report or the ISCA hardware paper as if either documents V3.1. Use a primary source that actually covers V3.1 for any version-specific technical or benchmark claim.
What this means for production API selection
A paper describes research claims; current API documentation describes the models and terms a provider documents for its service. Do not infer API availability, model behavior, limits, or price from a V3 or V3.1 paper.
As of November 10, 2026, DeepSeek’s official API pricing documentation lists deepseek-flash and deepseek-v4-pro. It maps deepseek-flash to DeepSeek-V4.1-Flash-1017. The same documentation says the legacy identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but their corresponding models have been retired; requests using them are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. Prices are per million tokens. For Flash, the page lists $0.15 per million input tokens on a cache miss and $0.003 on a cache hit off-peak, and $0.60 per million output tokens off-peak. Peak rates are $0.30 for cache-miss input, $0.006 for cache-hit input, and $1.20 for output. The docs define peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; other times are off-peak. Prices can change, so consult the live page before making a purchasing decision. The listed models do not establish whether V3.1 is available through the API.
Before choosing a production model:
- Confirm the served model ID in the provider’s current model-list documentation; don’t assume an older model name still maps to the same version.
- Check the documented capabilities and limits that matter to your application, such as context capacity, output limits, supported inputs, or tool calling.
- Test representative requests from your own workload. Compare answer quality, latency, and token cost under your expected input and output sizes. Paper architecture numbers alone cannot settle those operational trade-offs.
Bottom line: for V3 architecture, start with the DeepSeek-V3 Technical Report; for V3 hardware co-design, read Insights into DeepSeek-V3. If you mean a V3.1 paper, keep it distinct from these V3 publications and use a primary source that covers V3.1. For API selection, check the current official model list and pricing rather than treating a paper or release announcement as an API contract.
Sources
Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures