DeepSeek V4.1 Flash 1017: A “Minor” Update?
Read the Portuguese translation →
Author: Prompt Engineering · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: DeepSeek V4.1 Flash 1017: A “Minor” Update? · Channel: Prompt Engineering · 01:02 · Watch on YouTube
Topics: AI evaluation and benchmarks, Model inference and optimization, AI models, RAG, search and information retrieval
Summary
Section titled “Summary”DeepSeek has released a new version of its Flash model, which it describes as a minor update, though the video highlights substantial benchmark gains. The model reportedly improved on every benchmark in DeepSeek’s table, including a jump on Terminal-Bench 4.0 from 31% to 42% in five weeks. Its performance relative to price places it at the Pareto frontier, and the video compares it with GPT 5.6 while noting its much lower output price. DeepSeek says it used reinforcement learning on complete-project building, is launching a low-cost search API, and has made the model weights available on Hugging Face.
Key points
Section titled “Key points”- DeepSeek describes the Flash release as a minor update, but the video says its benchmark results show notable gains.
- The model improved on every benchmark in DeepSeek’s table.
- On Terminal-Bench 4.0, the score rose from 31% to 42% in five weeks.
- DeepSeek’s chart places the model at the Pareto frontier at about 20 cents per task.
- The video says its performance is comparable to OpenAI’s GPT 5.6, while its output price is about one-sixteenth as much.
- DeepSeek says it trained the model with reinforcement learning on building complete projects.
- DeepSeek is releasing a search API with 1,000 requests priced at 75 cents.
- The presenter reports getting about a third fewer thinking tokens in their own tests.
- The model is the same size as before, and its weights are available on Hugging Face.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Benchmark gains and pricing | The video introduces the Flash update and highlights its benchmark improvements and reported cost per task. |
| 00:31 | Comparisons, training, and search API | The presenter compares the model’s price and performance, describes its training approach, and introduces DeepSeek’s search API. |
| 01:02 | Concise responses and open weights | The presenter notes fewer thinking tokens in personal tests and closes by emphasizing the model’s optimization and open weights. |
Quotes
Section titled “Quotes”On terminal bench 4.0, it jumped from 31 to 42% in just 5 weeks — 00:00
it gets the highest score on the whole chart for about 20 cents a task — 00:00
Deep Seek says they trained it with reinforcement learning on building complete projects. — 00:31
In my own tests, I get about a third fewer thinking tokens. — 01:02
the weights are open, available on Hugging Face. — 01:02
Concepts and entities
Section titled “Concepts and entities”DeepSeek V4.1 Flash 1017, Terminal-Bench 4.0, DeepSeek, GPT-5.6, OpenAI, Hugging Face, reinforcement learning, Pareto frontier
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Deepseek just released their latest version of the flash model. They’re calling it a minor update, but the benchmarks tell a very different story. Here is all you need to know. It went on every benchmark in their table. On terminal bench 4.0, it jumped from 31 to 42% in just 5 weeks, which is kind of crazy. Now, the most interesting part is the pricing. On Deep Seek chart, it gets the highest score on the whole chart for about 20 cents a task, which puts it right at the Pareto frontier. It’s actually
00:31 comparable to Open AI’s GPT 5.6 sold. Now, that is not Open AI’s latest model, but this is a flash model at about a 16th of the output price, which is kind of crazy. Deep Seek says they trained it with reinforcement learning on building complete projects. And here are some of the demos that they shared, which looks pretty amazing. They’re also releasing their own search API, which you can get 1,000 requests for an incredible price of 75 cents only. They say the response is going to be
01:02 more concise with fewer tokens. In my own tests, I get about a third fewer thinking tokens. Deep Seek has always been about optimization and performance, and this is their latest Now, it’s the same model as before, same size, but they have pushed the Pareto frontier of optimization. The best part, the weights are open, available on Hugging Face.