DeepSeek V4.1 Flash 1017: A Table-Based Look At Agent Performance

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: A Table-Based Look At Agent Performance on ThorstenMeyerAI.com

DeepSeek released V4.1 Flash 1017 on October 17, replacing its September Flash model on the company’s website, app and API. DeepSeek says the update improved results on all 11 benchmarks in its model card and reduced the cost of completing agent tasks, while its published table also shows the model trailing competitors on several measures.

The new version is served under the API name deepseek-flash; DeepSeek says older aliases, including deepseek-v4-flash, now route to it as well. The listed price is unchanged. The weights are available under the MIT license, and the model accepts image input. DeepSeek describes the release as a minor update, but has not disclosed specific training changes.

In the company’s model-card table, Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench climbed from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, increased more modestly from 74.2 to 77.0. DeepSeek says the update also produces shorter outputs, which it presents as a way to finish tasks faster and reduce costs.

The table is not a clean measure of performance across every model: DeepSeek says Terminal-Bench was run in the Claude Code harness and DeepSWE in Mini-SWE-V2. It also notes that its HLE score uses the full set, while one rival’s displayed score is text-only. Across the 11 rows, 1017 leads four; Claude Opus 5 leads six and GPT-5.6 Sol leads one.

Lower Costs for Agent Work

The release matters less as a claim of overall model leadership than as a possible lower-cost option for repeated agent tasks. DeepSeek’s published prices are $0.15 per million uncached input tokens off-peak and $0.60 per million output tokens, with higher peak rates. The company’s schedule places peak hours on weekdays from 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays.

Those rates can affect teams that run coding or automation agents at scale, where repeated model calls add up. DeepSeek’s examples and benchmark results are company-reported, however, and do not by themselves establish what a particular customer will spend or how the model will perform on its own tasks. A small Kingy AI pilot, described by The Decoder, also points to a possible cost advantage but is too limited to settle broad comparisons.

From September Model to 1017

DeepSeek released the earlier V4.1 Flash on September 10. The October 17 version replaces it in the company’s hosted services, so API users calling the moving alias may receive 1017 without changing the model name. The model card’s comparison places the two DeepSeek versions alongside four competitors selected by DeepSeek.

According to the technical report as summarized by Kingy AI, the model uses a sparse Mixture of Experts architecture and a one-million-token context window. The report describes a Causal Encoder-Decoder design with 40 layers and 384 routed experts, six active per token, plus a shared expert. DeepSeek lists an FP8 checkpoint with FP4 experts. The technical report and reported architecture details provide scale, but they do not explain which changes produced the benchmark gains.

DeepSeek also says it tuned 1017 for its own DeepSeek Harness, recommending version 0.2.1-rc.1 or later, and for Tencent’s CodeBuddy and WorkBuddy. That focus on agent software helps explain why the largest listed improvements are on long, multi-step tasks rather than every type of evaluation.

„minor update“

— DeepSeek, in its model announcement

Independent Testing Still Limited

The benchmark figures in the model card come from DeepSeek and have not been independently reproduced in the material provided. The comparison also uses rivals that are a generation older than the latest named models. Kingy AI reported separately published Terminal-Bench results of 57 to 66 for current frontier models, but those evaluations used different conditions and are not a direct comparison.

There is also limited evidence about real-world reliability, speed and cost across varied workloads. The Kingy AI pilot covered seven small tasks, each run twice; all three tested models received perfect scores, and one 1017 task took about 27 seconds compared with about three seconds for Sonnet. DeepSeek has not specified what changed in training, and the available source material cuts off during a description of its newly opened search tools, leaving their full scope and launch details unclear.

More Testing and Tool Details

The next useful evidence will be independent evaluations that compare 1017 with current models using the same benchmarks, harnesses and task settings. Developers can also test it through the hosted API or the open-weight checkpoint, checking whether DeepSeek’s reported gains and price advantage hold for their own workloads.

Further details are still needed on the model’s training changes and on the search tools mentioned in the source report. Until those details and broader evaluations are available, the model card is the main basis for judging the release, and its results should be read as DeepSeek’s reported measurements, not independently established rankings.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026, replacing the September V4.1 Flash version on its website, app and API.

Did DeepSeek change the API price?

No price change was reported. DeepSeek says the update is served under deepseek-flash at the existing price, with separate peak and off-peak rates.

Which benchmarks showed the biggest gains?

In DeepSeek’s table, Terminal-Bench 4.0 rose by 11.1 points, from 31.2 to 42.3, and ProgramBench rose by 8.8 points, from 20.3 to 29.1.

Has an independent group confirmed DeepSeek’s benchmark results?

The source material says the model-card figures have not been independently reproduced. A small Kingy AI pilot reported results on seven tasks, but it is not enough to verify broad benchmark claims or rank the tested models.

Is 1017 the leading model in DeepSeek’s comparison?

No. It leads four of the 11 rows in DeepSeek’s table. Claude Opus 5 leads six, and GPT-5.6 Sol leads one; the comparison also does not include some newer models named in the source report.

Source: ThorstenMeyerAI.com