DeepSeek V4.1 Flash 1017: The Table To Read Before Changing Models

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: The Table To Read Before Changing Models on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, replacing the September version on its website, app and API while keeping the same listed price. The company reports higher scores on all 11 benchmarks it published, with its largest gains on some agent and coding tasks; the results are company-reported and have not been independently reproduced.

The update is served through the API name deepseek-flash; older names, including deepseek-v4-flash and deepseek-v4-flash-vision-exp, are also routed to the new version, according to the source report. DeepSeek describes the release as a minor update. The weights are available under the MIT license, and the model accepts image input as well as text.

In DeepSeek’s published table, Terminal-Bench 4.0 rises from 31.2 to 42.3, a gain of 11.1 points, while ProgramBench increases from 20.3 to 29.1. The gain on DeepSWE v1.1 is smaller, moving from 74.2 to 77.0. The table covers 11 tests, but 1017 leads four rows; Claude Opus 5 leads six and GPT-5.6 Sol leads one. These are scores on named evaluations, not a general measure of model quality.

DeepSeek also says 1017 produces shorter outputs, which can reduce completion time and token charges. The model card’s figures put API output at $0.60 per million tokens off-peak and $1.20 at peak, with input pricing varying by cache status and time. According to The Decoder’s account of DeepSeek’s cost chart, the model reached 77% on DeepSWE at just under $0.20 per task; that cost comparison is based on the cited evaluation, not a universal price for all coding work.

Lower Costs for Agent Tasks

The release matters most to developers running repeated, multi-step tasks through an API. If shorter outputs and the reported task costs hold in their own workloads, teams could test agent workflows at a lower price than with more expensive models. DeepSeek’s listed prices and the claimed benchmark gains make 1017 worth evaluating for cost-sensitive coding and automation jobs.

That does not establish that it is the best model for those jobs. Benchmark rank, task cost and reliability are separate questions, and results can vary with the task, tool harness, prompt and evaluation setup. DeepSeek’s table reports Terminal-Bench in the Claude Code harness and DeepSWE in Mini-SWE-V2. The source material also notes that figures for newer frontier models on Terminal-Bench come from separate evaluations, so they should not be treated as directly comparable.

An early Kingy AI pilot offers a limited outside check: across seven small tasks run twice, 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs. Kingy AI reported that 1017 cost roughly a quarter as much as the other two in that pilot, but it did not distinguish their quality because all three scored perfectly. One task took 1017 about 27 seconds, compared with about three seconds for Sonnet. The small test cannot settle broader questions of performance or speed.

What DeepSeek Changed

DeepSeek’s earlier V4.1 Flash version was released on September 10, 2026. Version 1017 replaces it on DeepSeek’s services, so users calling the general API alias receive the newer model without changing that alias. The supplied source says the release’s weights are hosted on Hugging Face and ModelScope as an FP8 checkpoint with FP4 experts.

The technical details in the source describe a sparse Mixture of Experts model with a reported one-million-token context window and up to 384,000 output tokens on the hosted API. A technical report, as summarized by Kingy AI, describes a Causal Encoder-Decoder design with 40 layers and 384 routed experts, six active per token plus a shared expert. Model-size figures depend on counting method: the report gives a 552-billion-parameter backbone, while Hugging Face lists 763 billion. Those figures should not be read as competing measurements of active parameters.

DeepSeek has not specified what changed in training. The Decoder attributes the gains to post-training with reinforcement learning and model self-feedback, but that is a report about the likely method, not a training explanation confirmed in the supplied material by DeepSeek. DeepSeek says it tuned 1017 for its own agent harness and names Tencent’s CodeBuddy and WorkBuddy as supported environments.

““minor update””

— DeepSeek, as characterized in the release material

Independent Tests Still Needed

The key uncertainty is whether DeepSeek’s reported gains and cost advantages carry over to independent evaluations and real deployments. The supplied source states that none of the 11 benchmark results has been independently reproduced. Its comparison table also uses rivals that the source describes as older-generation models, rather than the latest named releases.

It is also unclear what changed in the training process, how broadly the model’s shorter-output claim applies, and how performance varies across different agent harnesses and task types. The Kingy AI pilot covered only seven small tasks, and the source does not provide enough detail to generalize its results to larger projects. The table’s scores also mix evaluations with different tools and, in one case, different subsets: DeepSeek’s HLE full-set score is 42.6, while its text-only score is 44.0.

Test Against Your Workload

Developers can now try 1017 through DeepSeek’s website, app, API or released weights, with the API alias automatically pointing to the new version. The practical next step is to compare it with existing models on representative tasks, measuring completion quality, retries, latency and total cost rather than relying on a single benchmark score.

Independent evaluations would help clarify the reported gains, particularly on long agent runs, and allow more direct comparisons with current rival models. Until those tests are available, the October update is a lower-priced model release with promising company-reported results, not a settled ranking of the models available to developers.

Key Questions

What changed with DeepSeek V4.1 Flash 1017?

DeepSeek released 1017 on October 17, 2026, replacing the September V4.1 Flash version across its website, app and API. The company reports gains on all 11 benchmarks in its model card and says outputs are shorter. It kept the listed API price unchanged.

Is 1017 independently verified as better?

No. The supplied source says DeepSeek’s benchmark results have not been independently reproduced. Kingy AI reported a small pilot, but its three tested models tied across the tasks, so it did not establish that 1017 was better.

How does 1017 compare with the models in DeepSeek’s table?

It leads four of the 11 listed benchmark rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. The source also says the comparison omits newer rival versions, and separately reported scores are not always directly comparable.

What does DeepSeek charge for the API?

The source lists off-peak pricing of $0.15 per million input tokens for a cache miss, $0.003 for a cache hit and $0.60 per million output tokens. Peak prices are higher; the exact cost of a task depends on token use and cache status.

Can developers use the model weights?

Yes. The source says the weights are available under the MIT license on Hugging Face and ModelScope. DeepSeek also serves the model through its website, app and API.

Source: ThorstenMeyerAI.com