The answer in brief
- As of 31 October 2026, DeepSeek’s newest model is DeepSeek-V4.1-Flash-1017, released on 17 October. It runs DeepSeek’s website and app, and developers call it through the API as
deepseek-flash. - It is a 552-billion-parameter mixture-of-experts model that reads images as well as text, with a one-million-token context window and open weights under the MIT licence.
- It outscores DeepSeek’s larger V4 Pro. On version 5 of Artificial Analysis’s Intelligence Index it scores 32 to V4 Pro’s 23, level with GLM-5.3 for the top score among models with downloadable weights. The best proprietary models score up to 59.
- At peak times the API charges $0.30 per million input tokens and $1.20 per million output tokens, against $1.32 and $3.96 for V4 Pro. Both rates halve off-peak, which in winter covers UK weekdays from 10am and all weekend.
- DeepSeek has promised a larger V4.1 Pro but has not said when it will arrive.
- DeepSeek’s app and API store what you type in China. For client data, run the open weights on servers you control or through a host whose terms you have checked.
What are the latest DeepSeek models?
DeepSeek’s current line-up is two models: V4.1 Flash, in the update released on 17 October 2026, and V4 Pro, the 1.6-trillion-parameter flagship it previewed in April and released in full in August. Both are mixture-of-experts models, so only a small share of their parameters works on each token. Both have open weights under the MIT licence.
On the API, that comes down to two names. deepseek-flash serves the 1017 model and deepseek-v4-pro serves DeepSeek-V4-Pro-0813. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names still work, but DeepSeek answers them with the 1017 model and bills them at the Flash price.
If the table extends beyond the screen, scroll sideways to view all columns.
| Detail | DeepSeek V4.1 Flash | DeepSeek V4 Pro |
|---|---|---|
| Current version | DeepSeek V4.1 FlashDeepSeek-V4.1-Flash-1017 | DeepSeek V4 ProDeepSeek-V4-Pro-0813 |
| API model name | DeepSeek V4.1 Flashdeepseek-flash |
DeepSeek V4 Prodeepseek-v4-pro |
| Total parameters | DeepSeek V4.1 Flash552B, plus a 196B lookup memory | DeepSeek V4 Pro1.6T |
| Active parameters per token | DeepSeek V4.1 Flash8B reading input, 16B writing output | DeepSeek V4 Pro49B |
| Context window | DeepSeek V4.1 Flash1M tokens | DeepSeek V4 Pro1M tokens |
| Maximum output on the API | DeepSeek V4.1 Flash384K tokens | DeepSeek V4 Pro384K tokens |
| Image input | DeepSeek V4.1 FlashYes | DeepSeek V4 ProNo |
| Concurrent requests on the API | DeepSeek V4.1 Flash2,500 | DeepSeek V4 Pro500 |
| Open weights | DeepSeek V4.1 Flash510 GB, MIT licence | DeepSeek V4 Pro893 GB, MIT licence |
Sources: DeepSeek Models & Pricing, DeepSeek’s V4 preview post and the Hugging Face model cards and file listings, checked 31 October 2026.
The two names hide six model releases in six months. DeepSeek has also retired its old API names and published the engine that runs its own service.
If the table extends beyond the screen, scroll sideways to view all columns.
| Date | Release | What it changed |
|---|---|---|
| 20 October 2026 | ReleaseDeepInfer | What it changedOpen-sourced the inference engine behind DeepSeek’s own service |
| 17 October 2026 | ReleaseDeepSeek-V4.1-Flash-1017 | What it changedHigher scores and shorter outputs; Search and Fetch APIs opened |
| 10 September 2026 | ReleaseDeepSeek-V4.1-Flash | What it changedNew architecture with image input; replaced V4 Flash; lower Flash prices |
| 21 August 2026 | ReleaseDeepSeek-V4-Flash-Vision-Exp | What it changedExperimental image input on the API |
| 13 August 2026 | ReleaseDeepSeek-V4-Pro-0813 | What it changedV4 Pro left preview; peak and off-peak prices from 16 August |
| 31 July 2026 | ReleaseDeepSeek-V4-Flash-0731 | What it changedFinal V4 Flash version |
| 24 July 2026 | ReleaseOld API names retired | What it changeddeepseek-chat and deepseek-reasoner stopped working at 15:59 UTC |
| 24 April 2026 | ReleaseDeepSeek-V4 preview | What it changedV4 Pro (1.6T total, 49B active) and V4 Flash (284B, 13B), both with 1M context |
Sources: DeepSeek API news pages, Hugging Face and GitHub, checked 31 October 2026.
What changed in DeepSeek V4.1 Flash 1017?
The 17 October update scores higher than the September release on every test in DeepSeek’s table, while finishing tasks faster and writing shorter answers. DeepSeek calls it a minor update. The API name, calling method and prices did not change, so anyone already using deepseek-flash got the new model without touching their code.
DeepSeek trained the update with reinforcement learning that uses the model’s own feedback, aimed at building whole projects from scratch. Its examples include a web racing game and a 3D model of the Forbidden City. It tuned the model for DeepSeek Harness, its own agent framework, and for Tencent’s CodeBuddy and WorkBuddy, and recommends Harness 0.2.1-rc.1 or later.
If the table extends beyond the screen, scroll sideways to view all columns.
| Benchmark | V4.1 Flash 1017 | V4.1 Flash (September) | Claude Opus 5 | GPT-5.6 Sol | Kimi K3 | GLM-5.3 |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | V4.1 Flash 101777.0 | V4.1 Flash (September)74.2 | Claude Opus 574.0 | GPT-5.6 Sol73.0 | Kimi K367.5 | GLM-5.366.9 |
| Terminal-Bench 4.0 | V4.1 Flash 101742.3 | V4.1 Flash (September)31.2 | Claude Opus 551.8 | GPT-5.6 Sol39.9 | Kimi K312.6 | GLM-5.337.9 |
| ProgramBench | V4.1 Flash 101729.1 | V4.1 Flash (September)20.3 | Claude Opus 537.0 | GPT-5.6 Sol23.0 | Kimi K317.5 | GLM-5.319.0 |
| NL2Repo-Bench | V4.1 Flash 101768.2 | V4.1 Flash (September)65.4 | Claude Opus 575.3 | GPT-5.6 Sol56.8 | Kimi K358.0 | GLM-5.358.0 |
| Automation-Bench v1.0.6 | V4.1 Flash 101760.5 | V4.1 Flash (September)54.8 | Claude Opus 550.3 | GPT-5.6 Sol45.8 | Kimi K346.7 | GLM-5.348.8 |
| Agents’ Last Exam | V4.1 Flash 101735.4 | V4.1 Flash (September)31.8 | Claude Opus 528.6 | GPT-5.6 Sol26.7 | Kimi K327.6 | GLM-5.328.5 |
| Apex Bench v1.1 | V4.1 Flash 101772.1 | V4.1 Flash (September)64.6 | Claude Opus 565.8 | GPT-5.6 Sol54.3 | Kimi K347.8 | GLM-5.356.6 |
| Humanity’s Last Exam, no tools | V4.1 Flash 101742.6 (text-only subset 44.0) | V4.1 Flash (September)36.8 (text-only subset 39.1) | Claude Opus 556.3 | GPT-5.6 Sol44.5 | Kimi K343.5 | GLM-5.342.0 (text-only subset) |
DeepSeek’s own tests; higher is better. Source: DeepSeek-V4.1-Flash-1017 model card, checked 31 October 2026.
Read the rival columns with care. DeepSeek compared its 17 October release with Claude Opus 5 and GPT-5.6 Sol, both of which had newer successors by then: Anthropic released Claude Opus 5.5 on 22 September and OpenAI launched GPT-6.1 Sol on 29 September. The harness matters too. DeepSeek ran Terminal-Bench 4.0 inside the Claude Code harness and DeepSWE inside Mini-SWE-V2. The card also reports three image tests, left out of the table; Claude Opus 5 edges the 1017 model on all three, by one to two points.
The same release opened DeepSeek’s in-house search engine to developers as two APIs, Search and Fetch, at $0.75 per 1,000 successful requests each. Calls made through DeepSeek Harness, Claude Code or Codex on DeepSeek’s API were free until 4pm UK time on 31 October 2026.
How does the V4.1 architecture cut costs?
V4.1 Flash does less work reading than writing, and it keeps far less memory for each token in its context. Both savings land where agents spend most, because an agent re-reads the same files, pages and tool results at every step.
DeepSeek calls the design a Causal Encoder-Decoder. The model has 40 layers, split into a 20-layer encoder and a 20-layer decoder, and the decoder builds its long-range memory from the encoder’s output. That lets it use 8B active parameters per token while reading a prompt and 16B while writing an answer. V4 Flash used 13B for both.
The memory saving is bigger. The model card puts the global KV cache, the stored working memory for every token in context, at 890 bytes per token, about a quarter of V4 Flash. The copy kept on SSD shrinks to about an eighth. DeepSeek’s launch post says cache-hit charges often make up a large share of agent costs, and that compressing the cache cuts them.
The network underneath is still a mixture of experts. Each expert layer holds 384 routed experts and one shared expert, and six routed experts work on each token. Beside the main network sits Engram, a 196B-parameter memory the model looks up by token. Together they account for most of the 763B parameters Hugging Face counts in the checkpoint. A vision encoder trained from scratch fed images in from the start of pre-training, which ran on 45 trillion tokens.
The open weights take a reasoning effort from 1 to 100, which trades cost against accuracy. The API maps requests to three levels, low, high and max, and uses high unless told otherwise. The Artificial Analysis scores in this post for DeepSeek’s models are at max.
Is DeepSeek V4 Pro still available?
Yes. deepseek-v4-pro still runs on DeepSeek’s API at its own prices, after DeepSeek dropped a plan to fold it into V4.1 Flash.
When it launched V4.1 Flash on 10 September, DeepSeek said tests by several parties put the new model ahead of V4 Pro on performance, cost, speed and total runtime, and that it was phasing V4 Pro out. From 04:00 UTC on 14 September, V4.1 Flash was to answer every V4 Pro request at Flash rates until V4.1 Pro launched. By 11 September DeepSeek had changed course “in response to user demand”, keeping V4 Pro running with billing unchanged.
There are still reasons to pay for it. In DeepSeek’s tests of the base models, V4 Pro scores 55.2 on SimpleQA-Verified, a test of factual recall, against 42.3 for V4.1 Flash. On LongBench-V2, which tests long documents, it leads 51.5 to 45.2. Both are scores for the base models, before post-training, so they show what pre-training gave each model rather than how the API versions answer.
Against that, V4 Pro has no image input, a fifth of Flash’s concurrency limit, and it costs 4.4 times as much per input token and 3.3 times as much per output token. On version 5 of Artificial Analysis’s index it scores 23 to Flash’s 32. DeepSeek has promised V4.1 Pro as its successor but has not given a date.
How good is DeepSeek V4.1 Flash in independent tests?
It ties for the top score among models you can download, and sits far below the best proprietary models. On Artificial Analysis‘s Intelligence Index the 1017 model scores 32, level with GLM-5.3, while the leading proprietary models score up to 59.
Artificial Analysis moved to version 5 of the index at the end of October 2026. Most of its weight now rests on private test sets, and open-weight models lost the most ground. The 1017 model scored 45 on the old version 4.3.2, so figures quoted from the two versions cannot be compared. The announcement names Kimi K3.1 Pro as the top open-weights model, at 46, but Moonshot has not released its weights yet.
If the table extends beyond the screen, scroll sideways to view all columns.
| Model and reasoning effort | Weights | Intelligence Index v5 | Cost per index task |
|---|---|---|---|
| DeepSeek V4.1 Flash 1017 (max) | WeightsDownloadable | Intelligence Index v532 | Cost per index task$0.60 |
| GLM-5.3 (max) | WeightsDownloadable | Intelligence Index v532 | Cost per index task$5.01 |
| MiMo-V2.6-Pro | WeightsDownloadable | Intelligence Index v530 | Cost per index task$0.26 |
| DeepSeek V4 Pro 0813 (max) | WeightsDownloadable | Intelligence Index v523 | Cost per index task$1.77 |
| GPT-6.1 Sol (high) | WeightsProprietary | Intelligence Index v550 | Cost per index task$0.59 |
| GPT-6.1 Sol (max) | WeightsProprietary | Intelligence Index v555 | Cost per index task$1.27 |
Selected models. Source: Artificial Analysis Intelligence Index v5, at the list prices it records, checked 31 October 2026.
Against DeepSeek’s own range the result is clear: Flash beats V4 Pro by nine points for about a third of the cost per task. Artificial Analysis has not published a version 5 score for the September release, so the index cannot yet show how much the update itself added. Our Muse Spark Open test compares the wider open-weight field on the older version 4.3.2.
The comparison with proprietary models is less flattering. Artificial Analysis prices each task at list rates, and at DeepSeek’s peak rate the 1017 model costs the same per task as GPT-6.1 Sol at high effort, which scores 18 points more. DeepSeek’s off-peak rate halves its cost to about $0.30 a task for the same token use, and many hosts on OpenRouter charge less than DeepSeek’s peak rate around the clock.
Two cautions apply. The model talks a lot: Artificial Analysis ranks it 21st of 22 comparable models for verbosity, at 170M output tokens to run the index, and output is the expensive side of the bill. The agent harness also moves the score. DeepSeek’s own card for the September model shows DeepSWE results from 65.5% in OpenCode to 69.8% in Claude Code and 74.2% in Mini-SWE, so test it inside the agent you will actually run.
What does the DeepSeek API cost?
deepseek-flash costs $0.30 per million input tokens and $1.20 per million output tokens at peak times, and half that off-peak. deepseek-v4-pro costs $1.32 and $3.96 at peak. These are US dollar list prices, and DeepSeek’s pricing page does not mention VAT.
If the table extends beyond the screen, scroll sideways to view all columns.
| Model and period | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
deepseek-flash, peak |
Input, cache hit$0.006 | Input, cache miss$0.30 | Output$1.20 |
deepseek-flash, off-peak |
Input, cache hit$0.003 | Input, cache miss$0.15 | Output$0.60 |
deepseek-v4-pro, peak |
Input, cache hit$0.044 | Input, cache miss$1.32 | Output$3.96 |
deepseek-v4-pro, off-peak |
Input, cache hit$0.022 | Input, cache miss$0.66 | Output$1.98 |
Source: DeepSeek Models & Pricing, checked 31 October 2026.
Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday, except on Chinese public holidays. Every other hour is off-peak, weekends included. The UK has been on GMT since 25 October, so until British Summer Time returns on 28 March 2027 those are UK clock times too: weekday requests after 10am are billed at the off-peak rate until 1am. In summer the windows move an hour later on UK clocks.
Input that matches a prefix DeepSeek has already cached costs the cache-hit rate, which on Flash is a fiftieth of the normal input price. An agent re-sends its whole context at every step, so the cache-hit rate often decides the bill. DeepSeek runs its cache on a best-effort basis with no guaranteed hit rate, and each API response reports how many input tokens hit it.
The Flash price has moved twice this year. The 10 September cut took the August peak rates from $0.44 to $0.30 per million input tokens and from $1.32 to $1.20 for output. Even so, Flash costs more than V4 Flash did before 16 August, when DeepSeek moved to peak and off-peak pricing: the archived pricing page from July shows $0.14 and $0.28.
Can you run DeepSeek’s models yourself?
Yes, if you have data-centre hardware. The weights are free to download under the MIT licence, which allows commercial use, but the 1017 checkpoint alone is 510 GB and V4 Pro 0813 is 893 GB, before any memory for context.
DeepSeek publishes the weights on Hugging Face and ModelScope, with instructions for vLLM and SGLang. The checkpoints ship without a Jinja chat template. Instead DeepSeek supplies a Python reference encoder and deepseek-recipe, Rust libraries with Python bindings that turn API-style requests into the model’s prompt format.
On 20 October DeepSeek released DeepInfer under the Apache 2.0 licence. It is the inference engine behind DeepSeek’s own service, and it supports Nvidia’s SM100 GPUs and Huawei’s Ascend 950 chips. The open release covers offline inference only. It leaves out the KV cache system, the split between prefill and decode servers, and the load balancing that turn an engine into a public service. DeepSeek’s launch post for V4.1 Flash invites teams planning deployments of 2,000 GPUs with a storage cluster to get in touch, which gives a sense of the scale it has in mind.
For most UK businesses the practical route is a hosted copy of the open weights. Hugging Face lists Novita as an inference provider for the 1017 model, and your data then goes to that host under its terms. Check who actually serves the request: OpenRouter spreads requests for the 1017 model across more than 30 providers, DeepSeek among them, so a request sent that way can still reach DeepSeek unless you limit it to named providers.
Is DeepSeek safe for UK businesses to use?
It depends which DeepSeek you use. The app, website and API send what you type to DeepSeek in China, while the open weights run wherever you put them, and a suitably configured deployment keeps your prompts away from DeepSeek.
DeepSeek’s privacy policy, last updated on 10 February 2026, says: “we directly collect, process and store your Personal Data in People’s Republic of China.” It also lists training and improving its models among the uses of that data, and a right to opt out of that use by emailing DeepSeek. The policy does not cover tools other developers build on DeepSeek’s API. That developer is the controller and is responsible for telling its users how it handles their data, but your text still reaches DeepSeek.
China has no UK adequacy regulations, so sending personal data there is a restricted transfer. It needs appropriate safeguards, with a transfer risk assessment, or one of the limited exceptions.
The UK regulator has dealt with DeepSeek directly. On 8 October 2026 the ICO, which became the Information Commission on 30 September but kept the ICO name, named ten foundation model developers that had made, or committed to make, data protection changes after its scrutiny: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. The ICO says it is monitoring their progress against those commitments.
US rules are tighter. A Department of Homeland Security order, published in the Federal Register on 20 October, bars DeepSeek and products developed using it from US federal civilian agencies. Our guide to the DHS order covers who it reaches, and UK firms that sell software to US agencies should check whether anything they supply is built on DeepSeek.
Our view: keep client and customer personal data out of DeepSeek’s app and API. If the model suits the work, run the open weights on infrastructure you control, or through a host whose data location and terms you have checked, and record that choice in your data protection impact assessment. For material with no personal data or confidential business information in it, such as drafting from published sources or testing open-source code, the hosted API is a sound choice on price.
Which DeepSeek model should you use?
Of DeepSeek’s two models, V4.1 Flash, called as deepseek-flash, is the one to use for most work. It scores higher than V4 Pro on Artificial Analysis’s index, costs less, and is the only one of the two that reads images.
- Coding agents and automation. Use Flash, and schedule long runs for UK afternoons, evenings or weekends to pay the off-peak rate. Test it in the harness you will use, because scores shift by several points between harnesses. If your agents search the web through DeepSeek’s Search and Fetch APIs, budget $0.75 per 1,000 calls: the free period for DeepSeek Harness, Claude Code and Codex ended on 31 October.
- Screenshots, charts and scanned pages. Flash is the only choice, since V4 Pro has no image input.
- Questions that lean on the model’s own knowledge. V4 Pro’s base model leads on SimpleQA-Verified, the factual-recall test in DeepSeek’s results. Compare both on a sample of your own questions before paying over three times as much per token.
- Client or customer personal data. Use the open weights on your own servers or a host you have vetted.
One more practical point: deepseek-flash is a moving target. DeepSeek switched it from the September model to the 1017 update with no change on the caller’s side, and its API offers no dated name for the older version. If your outputs must stay stable, run a dated checkpoint yourself.
Price the work against a proprietary model before you switch to save money. On Artificial Analysis’s index, GPT-6.1 Sol at high effort costs the same per task as Flash at DeepSeek’s peak rate and scores 18 points more. Flash comes out cheaper off-peak or through a lower-priced host, though the cheapest hosts run the weights at reduced precision. Your own bill depends on how many tokens your work uses and how often the cache hits.
V4.1 Pro is the release to watch. DeepSeek has tied V4 Pro’s future to it but given no date. Kimi K3.1 Pro matters too: once Moonshot releases its weights, Flash no longer shares the top score among downloadable models.
If you are weighing DeepSeek against the models you already use, talk to AIWIZ.
Sources
- DeepSeek-V4.1-Flash-1017 release, DeepSeek API Docs, 17 October 2026
- DeepSeek-V4.1-Flash release, DeepSeek API Docs, 10 September 2026
- DeepSeek-V4-Pro general availability, DeepSeek API Docs, 13 August 2026
- DeepSeek V4 preview release, DeepSeek API Docs, 24 April 2026
- Models and pricing, DeepSeek API Docs
- DeepSeek-V4.1-Flash-1017 and DeepSeek-V4.1-Flash model cards, Hugging Face
- DeepInfer, DeepSeek on GitHub
- Announcing Artificial Analysis Intelligence Index v5, Artificial Analysis
- DeepSeek V4.1 Flash 1017 model page, Artificial Analysis
- DeepSeek privacy policy, last updated 10 February 2026
- ICO secures changes from leading AI developers as scrutiny extends to AI agents, ICO, 8 October 2026
- Information Commission succeeds the ICO as UK’s data protection regulator, GOV.UK, 30 September 2026
- DeepSeek V4 Pro API service will not be discontinued, PANews, 11 September 2026
- DeepSeek V4.1 Flash 1017, OpenRouter
- Is the restricted transfer covered by adequacy regulations?, ICO