Skip to content
FrontierDot

DeepSeek V4.1 Flash 1017: Release Timeline, Parameters, Prices and What Changed Since September

Live timeline · last updated · 1 entry in the update log

Developers calling deepseek-flash kept the same model name and the same price. What answered them changed on October 17, five weeks after V4.1 Flash first shipped.

DeepSeek released V4.1 Flash, an MIT-licensed open-weight MoE with a 552B-parameter backbone, on Sep 10, 2026, and replaced it on Oct 17 with V4.1-Flash-1017, which keeps the same size, API name and price and posts higher scores in DeepSeek’s own tests.

Read on 2026-10-19 (UTC+8): DeepSeek’s API change log, its Sep 10 and Oct 17 news pages, the Models & Pricing page, both Hugging Face model cards and their repository metadata, and official X posts by @deepseek_ai, @DeepSeekHarness, @ArtificialAnlys, @ValsAI, @ollama and @baseten. X times are UTC as X reports them. Scores are DeepSeek-reported unless a third party is named. We did not run the model.

Timeline

  1. DeepSeek’s API adds V4-Pro and V4-Flash under the model names deepseek-v4-pro and deepseek-v4-flash. The change log says the older names deepseek-chat and deepseek-reasoner will be discontinued in three months, on 2026-07-24, and that until then they point to the non-thinking and thinking modes of V4-Flash.

    DeepSeek API change log (2026-04-24)

  2. The official V4-Flash API goes into public beta. DeepSeek says this build, 0731, keeps the architecture and size of the preview and was only post-trained again, and it lists 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE. The note adds that V4-Pro and the app and web models were left unchanged.

    DeepSeek API change log (2026-07-31)

  3. V4-Pro reaches general availability on the app, web and API. The same entry adds three thinking-effort levels (low, high and max) and peak and off-peak pricing, with off-peak set at half the peak price from 16:00 UTC on August 16.

    DeepSeek API change log (2026-08-13)

  4. DeepSeek puts an experimental vision model, DeepSeek-V4-Flash-Vision-Exp, on the API. According to the change log its text abilities match V4-Flash, and on agent benchmarks that need visual understanding it comes close to Opus-4.8.

    DeepSeek API change log (2026-08-21)

  5. At 06:10 UTC DeepSeek opens a six-post thread on X announcing V4.1-Flash, which it calls the smallest model in a new architecture family, with native visual understanding. The thread describes a 552B-parameter MoE whose causal encoder-decoder design activates 8B parameters for input and 16B for output, and its last post links the weights and the technical report on Hugging Face.

    DeepSeek on X: Introducing DeepSeek-V4.1-Flash

  6. In the fourth post of that thread DeepSeek retires V4-Flash and V4-Flash-Vision-Exp, sends their old model names to V4.1-Flash, and says it is phasing out V4-Pro. From 04:00 UTC on September 14, it writes, every deepseek-v4-pro request will be routed to V4.1-Flash at V4.1-Flash rates until V4.1-Pro launches.

    DeepSeek on X: V4.1-Flash on the API

  7. Artificial Analysis, an independent benchmarking company, posts its results at 20:36 UTC. V4.1 Flash scores 40 on its Intelligence Index, ahead of DeepSeek V4 Pro 0813. The same post calls it one of the most verbose models the company has measured, at 89k tokens per index task, and still estimates its cost at $0.27 per task because the token prices are so low.

    Artificial Analysis on X

  8. Ollama posts at 23:12 UTC that V4.1-Flash is rolling out on its cloud, starting with Max and Team accounts, and that it is adding capacity to reach all subscribers.

    Ollama on X

  9. The DeepSeek Harness account announces desktop versions of DeepSeek’s agent harness for macOS and Windows, in an X article titled “DeepSeek Harness for Desktop Is Here”.

    DeepSeek Harness on X

  10. At 08:35 UTC DeepSeek announces DeepSeek-V4.1-Flash-1017, live on its website, app and API. The four-post thread claims stronger benchmark scores, faster task completion and shorter outputs than the September build, asks DeepSeek Harness users to upgrade to v0.2.1-rc.1 or later, and ends with links to the weights on Hugging Face and ModelScope.

    DeepSeek on X: DeepSeek-V4.1-Flash-1017 is here

  11. The third post of the same thread opens DeepSeek’s Search and Fetch APIs to everyone at $0.75 per 1,000 calls for each endpoint. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex stay free until billing starts at 16:00 UTC on October 31.

    DeepSeek on X: Search & Fetch APIs

Why the name confuses people

DeepSeek’s naming has two layers, and most of the confusion lives between them. One layer is the model version, a dated build such as V4-Flash-0731 or V4.1-Flash-1017. The other is the API model name you put in a request, which since September 10 has been deepseek-flash. When DeepSeek ships a new build it usually keeps the API name and changes what answers it. The change log entry for October 17 says so directly: the web and app models were updated, the calling method and pricing are unchanged, and deepseek-flash now reaches the new model.

A developer who set deepseek-flash in September has therefore been talking to 1017 since October 17 without changing a line of code. The older names still work. The Models & Pricing page, as we read it on October 19, says deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, that the models behind them have been retired, and that those requests are served by V4.1-Flash-1017 and billed at the Flash price.

The suffix is a date, October 17, written the same way as 0731 and 0813 for the earlier V4 builds. The 1017 model card calls it a minor update to V4.1-Flash. It sits in its own Hugging Face repository next to the September one, and both are released under the MIT License.

What the September architecture changed

The V4.1-Flash model card describes a Causal Encoder-Decoder: a 40-layer Transformer split into a 20-layer causal encoder followed by a 20-layer decoder. The decoder’s global KV cache is projected from the encoder’s final hidden states instead of being built in every decoder layer. DeepSeek says this is why only 8B parameters are active per token during prefill and 16B during decode, and says the design pays off most in input-heavy agent workloads.

The memory figures are what DeepSeek leads with. According to the card, a set of attention changes it calls CSA2, together with FP4 storage for the main KV cache, bring the global cache down to 890 bytes per token, about a quarter of V4-Flash. A second mechanism, SWA Bounded Replay, avoids writing sliding-window cache to SSD and shrinks the persistent cache to roughly an eighth of V4-Flash. A chart on the card puts the reduction since DeepSeek’s first-generation model at 437-fold. The announcement thread gives the commercial reason: cache-hit charges often account for a large share of what agents cost.

Other details on the card: each MoE layer has one shared expert and 384 routed experts, six of which are active per token; a 196B-parameter Engram conditional memory is read by token-based lookup; pre-training used 45T tokens of multimodal data, with context extended to 1M tokens after 34T of them. Reasoning effort is set as an integer from 1 to 100; the V4 models used three levels, low, high and max. Images pass through a vision encoder DeepSeek trained from scratch, and image and text data were mixed from the start of pre-training.

552B or 763B: reading the parameter count

Search for the size of this model and you get two numbers, both from DeepSeek’s own pages. The announcement and the base-model table say 552B, which the card labels backbone parameters. Hugging Face shows 763B for the repository. Most of the gap is the Engram memory, which the card puts at 196B parameters and which the model looks up instead of computing through on every token. Artificial Analysis gives both figures in a single post, calling it a 552B model in the opening line and listing 763B total parameters further down.

The download is easier to pin down. Hugging Face’s repository metadata lists 763,205,315,794 parameters and 510.30 GB of safetensors files, about 475 GiB. The 1017 repository shows the same parameter count and the same safetensors total, which is consistent with DeepSeek calling 1017 a minor update.

Weights of that size are aimed at servers, and DeepSeek’s announcement says as much: it invites teams planning a deployment of 2,000 GPUs plus a storage cluster to contact it, and says it is working with the open-source community on inference support. Hosted access was there from day one: Baseten and Ollama both announced V4.1-Flash on September 10, and the Hugging Face page lists Novita, Fireworks, Featherless AI, DeepInfra and Baseten as inference providers.

Prices, and the V4-Pro reversal

The September 10 pricing post said new prices would take effect at 04:00 UTC that day and that off-peak rates would stay at half the peak rate. On the pricing page as of October 19, deepseek-flash costs $0.30 per million input tokens on a cache miss, $0.006 on a cache hit and $1.20 per million output tokens at peak, and $0.15, $0.003 and $0.60 off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. The context window is 1M tokens and the maximum output is 384K.

The same page still lists deepseek-v4-pro, now pinned to DeepSeek-V4-Pro-0813, at $1.32 per million input tokens on a cache miss and $3.96 per million output tokens at peak. That runs against the plan in the September 10 thread, which was to send every V4-Pro request to V4.1-Flash from September 14. The change log accounts for it. Under its September 10 heading it now says that, in response to user demand, DeepSeek decided to keep serving V4 Pro after September 14 with billing unchanged. The entry carries no date of its own, and we found no DeepSeek post on X announcing the change.

Token prices do not settle the bill on their own, because a model that writes more tokens pays for more of them. Artificial Analysis measured V4.1 Flash at 89k tokens per task on its index, more than GLM-5.3, V4 Pro 0813 and frontier models such as Claude Opus 5. Its estimated cost per task, $0.27, still came out about 2.5 times below V4 Pro 0813. Vals AI, another independent evaluator, called it the top open-weight model on its index at $0.30 per test. Both are third-party results for the September build, not DeepSeek’s figures.

What 1017 changes, by DeepSeek’s numbers

The October 17 change log and the 1017 model card list the same eleven scores and set them against the September build. DeepSWE v1.1 goes from 74.2 to 77.0. Terminal-Bench 4.0 rises from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, Automation-Bench from 54.8 to 60.5, and HLE without tools from 36.8 to 42.6. The vision agent tests move less: BabyVision with tools from 89.6 to 92.5 and ZeroBench-main from 49.0 to 51.0.

Check the footnotes before you compare. For 1017, DeepSeek ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE in Mini-SWE-V2. The September card ran most code-agent tests in the Minimal mode of DeepSeek Harness, with mini-SWE for DeepSWE. Neither page says whether the September Terminal-Bench 4.0 score was rerun in Claude Code for the new table. DeepSeek’s pages also disagree on one September number: NL2Repo-Bench is 64.0 in the original model card and 65.4 in the change log and the 1017 table.

DeepSeek also says 1017 finishes tasks faster and writes shorter outputs, which would address the verbosity Artificial Analysis measured in September. Neither of DeepSeek’s pages gives a token count, so until someone measures it again, shorter output remains DeepSeek’s claim. The 1017 news page adds that the model was trained with reinforcement learning using self-feedback on end-to-end projects, and shows four demos, from a 3D model of the Forbidden City to a Portal 2 recreation built in Godot.

Two smaller differences show up in the repositories. The recommended top_p for 1017 is 1.0, where the September card allowed 0.95 or 1.0. The 1017 repository includes tokenizer files but not the technical report PDF, which remains in the September repository.

Search and Fetch, released the same day

The second announcement on October 17 is easy to miss. DeepSeek opened its Search and Fetch APIs, which it says run on its in-house search engine and handle hundreds of millions of requests a day. Each endpoint costs $0.75 per 1,000 successful requests. Calls that DeepSeek Harness, Claude Code or Codex make through server-side tools are free for now. The X thread puts the start of billing at 16:00 UTC on October 31, and the pricing page gives the same moment as 00:00 Beijing time on November 1.

Agent builders should put that date in the calendar. Server-side search calls happen whenever the agent decides to look something up, so a session that uses the same number of tokens can cost more after October 31 than it did the week before.

What we could not verify

  • When DeepSeek decided to keep serving V4-Pro after September 14. The change log records the decision under its September 10 heading without a date of its own.
  • Whether 1017 writes shorter outputs than the September build, and by how much. DeepSeek gives no token counts, and the Artificial Analysis verbosity figures are for the September build.
  • What hardware you need to run the weights yourself. DeepSeek publishes inference code and file sizes, but the pages we read give no minimum hardware list, and we did not run the model.
  • Third-party scores for 1017. The Artificial Analysis and Vals results cited here were posted on September 10 for the original V4.1-Flash.

Our take

V4.1 Flash is easy to follow line by line, because DeepSeek writes almost everything down: a dated change log, two model cards with benchmark tables, a pricing page with exact hours, and X threads with timestamps. The one soft spot is the V4-Pro decision, announced in an X thread and reversed in the change log without a post we could find. For anyone on the API the practical points are short. deepseek-flash already points to 1017, the price did not change, the old V4 names still resolve, and Search and Fetch billing starts at 16:00 UTC on October 31. We will add dated nodes here when DeepSeek ships V4.1-Pro, changes prices or publishes its next build.

Related reading on Front: the Ling 3.1 Flash hardware breakdown works through the same total-versus-active memory question for another large MoE, the Reflection Beam timeline and the Meta Muse Spark Open timeline follow two other open-weight releases from October, and Grok 4.8: what xAI has actually said tracks a model that has been talked about but not shipped. Every post is on the Front blog index.

The longer release line, from V2 through this build, is on the DeepSeek model release timeline.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

What is DeepSeek V4.1 Flash 1017?

DeepSeek-V4.1-Flash-1017 is the October 17, 2026 update to DeepSeek V4.1 Flash. DeepSeek’s model card calls it a minor update. It has the same parameter count and is served under the same API name, deepseek-flash, at the same price.

How many parameters does DeepSeek V4.1 Flash have?

DeepSeek gives 552B backbone parameters, with 8B active per token for input and 16B for output. Hugging Face lists 763B for the repository, which includes the 196B-parameter Engram memory. The safetensors files total about 510 GB.

How much does the DeepSeek V4.1 Flash API cost?

At peak hours, $0.30 per million input tokens on a cache miss, $0.006 on a cache hit and $1.20 per million output tokens. Off-peak prices are half that, according to DeepSeek’s pricing page as of October 19, 2026.

Is DeepSeek V4.1 Flash open source?

The weights and repository code for both V4.1-Flash and V4.1-Flash-1017 are published on Hugging Face under the MIT License, and DeepSeek also posted 1017 on ModelScope.

Update log

  • — First published. Checked DeepSeek’s pages and official X posts through Oct 19, 2026 (UTC+8).

Sources