DeepSeek V4 Release Date: From the April Preview to V4.1-Flash

DeepSeek V4 release date: the Preview shipped 24 April 2026, V4-Pro-0813 on 13 August and V4.1-Flash-1017 on 17 October. Timeline, prices and migration.

News·April 24, 2026·By DS Guide Editorial
Title card: eyebrow 'NEWS · V4 RELEASE TIMELINE', title 'DeepSeek V4 Releases', tagline '2026 release dates (month/day), per the Change Log.', and four stats: 4/24 V4 Preview; 8/13 V4-Pro-0813 GA; 9/10 V4.1-Flash; 10/17 V4.1-Flash-1017, beside a typewriter illustration.

The DeepSeek V4 release date depends on which V4 you mean. The V4 Preview dropped on April 24, 2026, as two open-weight Mixture-of-Experts models — deepseek-v4-pro and deepseek-v4-flash — both with a one-million-token context window. DeepSeek’s Change Log then dates the official releases that replaced it: DeepSeek-V4-Flash-0731 (API public beta, 31 July 2026), DeepSeek-V4-Pro-0813 (general availability, 13 August 2026), DeepSeek-V4.1-Flash as deepseek-flash (10 September 2026) and DeepSeek-V4.1-Flash-1017 (17 October 2026). After months of rumour (Hunter Alpha on OpenRouter, Reuters leaks about Huawei Ascend hardware, an “any day now” consensus on X), DeepSeek published the preview weights to Hugging Face, flipped the API over, and swapped the chat.deepseek.com default model in a single morning. This article gives you the verified timeline, what each V4 tier was at launch, current pricing, the April benchmark picture in context, and the migration path off the legacy deepseek-chat and deepseek-reasoner IDs, which were scheduled for retirement on July 24, 2026.

The confirmed DeepSeek V4 release date

Illustration of a developer at a laptop terminal sending a DeepSeek V4-Pro chat request and getting a JSON reply, with API docs on a second monitor

DeepSeek posted the V4 Preview announcement on its API docs site on April 24, 2026, the same morning the weights went live on Hugging Face and Expert/Instant Mode appeared in the web chat. The post opened by declaring DeepSeek-V4 Preview officially live and open-sourced, welcomed “the era of cost-effective 1M context length,” and introduced two preview models: DeepSeek-V4-Pro (Preview) at 1.6T total / 49B active parameters, and DeepSeek-V4-Flash (Preview) at 284B total / 13B active parameters.

Mainstream coverage locked in within hours. Bloomberg reported on April 24, 2026 that DeepSeek rolled out preview versions of its new flagship, unveiling the V4 Flash and V4 Pro series with top-tier coding-benchmark performance and advances in reasoning and agentic tasks. CNN’s coverage from the same day framed it as a preview rival to OpenAI, Anthropic and Google, with upgrades in reasoning, agentic ability, and token-processing efficiency.

The “Preview” label was not cosmetic: the April models have since been superseded. On 31 July 2026 DeepSeek put the official DeepSeek-V4-Flash-0731 into API public beta (same architecture and size as the preview, re-post-trained); on 13 August 2026 DeepSeek-V4-Pro-0813 reached general availability on app, web and API; on 10 September 2026 DeepSeek-V4.1-Flash launched as deepseek-flash with native image understanding, and V4 Flash was retired; on 17 October 2026 it was updated to DeepSeek-V4.1-Flash-1017. Today deepseek-flash is served by DeepSeek-V4.1-Flash-1017 and deepseek-v4-pro by DeepSeek-V4-Pro-0813 (source: DeepSeek’s Change Log, read October 2026).

What V4 actually is: two tiers, one feature set

At the April launch, V4 was not a single model — it was a family of two open-weight MoE checkpoints that shared architecture, context length, and the same three reasoning modes. You picked the tier with the model field and the effort with a parameter. The table below describes those April 2026 preview models; the parameter counts and disk sizes do not apply to today’s DeepSeek-V4-Pro-0813 or DeepSeek-V4.1-Flash-1017.

Attribute (April 2026 preview) DeepSeek-V4-Pro (Preview) DeepSeek-V4-Flash (Preview)
Total parameters 1.6T 284B
Active per token 49B 13B
Context window 1,000,000 tokens 1,000,000 tokens
Max output 384,000 tokens 384,000 tokens
Reasoning modes Non-thinking, high, max Non-thinking, high, max
Weights license MIT MIT
On-disk size (HF) ~865 GB ~160 GB
Release date 2026-04-24 2026-04-24
V4 Preview tier specifications, sourced from the DeepSeek API docs announcement and Hugging Face model card.

Simon Willison’s day-of writeup independently confirmed the preview’s sizing: Pro is 1.6T total with 49B active, Flash is 284B total with 13B active, both under the MIT license, with Pro at 865GB on Hugging Face and Flash at 160GB — making the V4-Pro preview, at the time, the largest open-weight model published, larger than Kimi K2.6 and more than twice the size of V3.2. For a deeper technical read on DeepSeek V4 as a family, see our dedicated model page; tier-specific deep dives live at DeepSeek V4-Pro and DeepSeek V4-Flash.

Architecture: why the 1M context is the actual story

Context length alone is marketing. What matters for agent workloads is how cheap each forward pass stays at that depth. DeepSeek’s April 2026 preview model card on Hugging Face describes a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA); in the 1M-token setting, DeepSeek-V4-Pro (Preview) requires only 27% of the single-token inference FLOPs and 10% of the KV cache compared with V3.2 (DeepSeek-reported). Both current models keep the 1M-token context, with up to 384K tokens of output.

Pricing — current rates and the off-peak status

The April preview launched with its own introductory rates, which no longer apply. DeepSeek introduced peak and off-peak pricing with the V4-Pro GA release (effective 16:00 UTC on 16 August 2026, off-peak at half the peak rate), and the 10 September 2026 V4.1-Flash release reduced API prices; the 17 October update left them unchanged. As of October 2026 the official pricing page lists the rates below, in USD per million tokens. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak. Requests that still use deepseek-v4-flash are served by DeepSeek-V4.1-Flash-1017 and billed at the deepseek-flash price.

Model (peak / off-peak) Input (cache hit) Input (cache miss) Output
deepseek-flash (DeepSeek-V4.1-Flash-1017) $0.006 / $0.003 $0.30 / $0.15 $1.20 / $0.60
deepseek-v4-pro (DeepSeek-V4-Pro-0813) $0.044 / $0.022 $1.32 / $0.66 $3.96 / $1.98
DeepSeek API pricing as of October 2026, USD per 1M tokens, peak / off-peak. Check the DeepSeek API pricing page for updates.

Pro costs 3.3× Flash on output and 4.4× on cache-miss input (at peak and off-peak alike), and only deepseek-flash accepts images, so Flash remains the default recommendation for most chat and standard coding work. DeepSeek’s 10 September 2026 news page says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, so price-check Pro against your own workload before choosing it. For live modelling, our DeepSeek pricing calculator handles the three-bucket math for you.

Worked example: 1M Flash calls with a cached system prompt

A common pattern: 1,000,000 calls, 2,000-token system prompt cached across calls, 200-token user message per call (uncached — each new user turn misses the prefix), 300-token response.

  • Cached input: 2,000 × 1,000,000 = 2.0B tokens × $0.006/M peak = $12.00 ($0.003/M off-peak = $6.00)
  • Uncached input: 200 × 1,000,000 = 200M tokens × $0.30/M peak = $60.00 ($0.15/M off-peak = $30.00)
  • Output: 300 × 1,000,000 = 300M tokens × $1.20/M peak = $360.00 ($0.60/M off-peak = $180.00)
  • Total on deepseek-flash: $432.00 at peak, $216.00 off-peak

The same workload at deepseek-v4-pro rates runs $88.00 + $264.00 + $1,188.00 = $1,540.00 at peak, or $44.00 + $132.00 + $594.00 = $770.00 off-peak (prices as of October 2026). Do not skip the uncached line — the user message on each call is still a miss against that cached prefix.

Benchmarks: competitive, not category-leading

The benchmark picture below is the April 2026 preview’s. DeepSeek was clear-eyed in its own announcement: it said V4-Pro (Preview) led all open models on world knowledge while trailing Gemini-3.1-Pro, and rivalled top closed-source models on reasoning rather than beating them. The preview model card’s table, for V4-Pro in its Max reasoning mode, shows the same mixed result. For the current releases DeepSeek reports different benchmark sets — for example HLE without / with tools 42.7 / 60.0 for DeepSeek-V4-Pro-0813 (13 Aug 2026 Change Log) and DeepSWE v1.1 77.0 for DeepSeek-V4.1-Flash-1017 (17 Oct 2026 Change Log) — so they are not comparable with the April numbers.

Benchmark V4-Pro Preview (Max) Claude Opus 4.6 GPT-5.4 Gemini 3.1 Pro
SWE-Bench Verified 80.6 80.8 n/r 80.6
Terminal-Bench 2.0 67.9 65.4 75.1 68.5
LiveCodeBench 93.5 88.8 — —
HMMT 2026 (math) 95.2 96.2 97.7 —
HLE (no tools) 37.7 40.0 39.8 44.4
SimpleQA-Verified 57.9 — — 75.6
DeepSeek-reported scores from the April 2026 V4-Pro preview model card (competitor columns: Opus-4.6 Max, GPT-5.4 xHigh, Gemini-3.1-Pro High, as DeepSeek labelled them). These do not describe DeepSeek-V4-Pro-0813.

OfficeChai’s April 2026 reporting on the preview summarised the same pattern: V4-Pro leads on IMOAnswerBench (89.8) against Claude (75.3) and Gemini (81.0) but trails GPT-5.4 (91.4); on HMMT 2026, Claude (96.2) and GPT-5.4 (97.7) pull decisively ahead; SWE-Verified shows V4-Pro at 80.6, within a fraction of Claude (80.8) and matching Gemini (80.6); Terminal-Bench 2.0 has V4-Pro (67.9) beating Claude (65.4) and close to Gemini (68.5), with GPT-5.4 leading at 75.1. For a fuller benchmark roundup, see our DeepSeek benchmarks 2026 page.

How to access V4 today

Three routes went live on release day.

  1. Web chat — DeepSeek’s 24 April 2026 announcement pointed users to chat.deepseek.com via Expert Mode / Instant Mode; its 13 August 2026 notes said V4 Pro was available on app and web via “Expert Mode”, and the 17 October 2026 Change Log says the official web and app models were updated to DeepSeek-V4.1-Flash-1017.
  2. API — POST /chat/completions against https://api.deepseek.com, the OpenAI-compatible endpoint. The Anthropic-compatible surface uses its own base URL, https://api.deepseek.com/anthropic.
  3. Open weights — published on Hugging Face under the DeepSeek collection.

A minimal Python snippet using the OpenAI SDK against V4-Pro with thinking mode enabled:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a careful engineer."},
        {"role": "user", "content": "Plan the migration from deepseek-chat."},
    ],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

print(resp.choices[0].message.content)

Thinking mode is the default on both current models, and reasoning_effort accepts low, high (the default) or max. When thinking is enabled the response returns reasoning_content alongside the final content. Core knobs worth knowing: temperature (non-thinking mode only, where DeepSeek recommends 0.0 for code, 1.3 for general chat, 1.5 for creative writing; thinking mode ignores it), top_p, max_tokens (set high enough to avoid JSON-mode truncation), stream=true, context caching (automatic on repeated prefixes), tool calling, FIM completion in Beta (non-thinking only: POST /completions on the Beta base URL, which takes no thinking field), and Chat Prefix Completion in Beta. If you are starting from scratch, the DeepSeek API getting started walkthrough covers account setup, keys, and the first request. JSON mode is designed to return valid JSON, not guaranteed — include the word “json” plus a schema example in your prompt.

Think Max: the context-window gotcha

DeepSeek’s April 2026 preview model card recommended setting the context window to at least 384K tokens for the Think Max reasoning mode when running the weights locally; the V4-Pro-0813 card likewise recommends a maximum output length of 384K tokens for the high and max effort levels. If you set reasoning_effort="max" on a local deployment without raising max_model_len, expect truncation.

Migration off legacy model IDs

If your integration still pointed at deepseek-chat or deepseek-reasoner, the deadline has passed. The April announcement stated that deepseek-chat and deepseek-reasoner would be fully retired and inaccessible after July 24, 2026, 15:59 UTC; until then they routed to deepseek-v4-flash in non-thinking and thinking modes respectively, and they are no longer listed in DeepSeek’s docs. DeepSeek’s guidance was to keep base_url and just update the model; since 10 September 2026 the Change Log’s instruction is to use deepseek-flash for the latest V4.1 Flash model (the legacy name deepseek-v4-flash is still accepted and served by DeepSeek-V4.1-Flash-1017).

  • deepseek-chat → deepseek-flash with extra_body={"thinking": {"type": "disabled"}} for non-thinking behaviour (thinking is the default).
  • deepseek-reasoner → deepseek-flash (thinking is on by default; reasoning_effort="high" is the default effort).
  • Want DeepSeek-V4-Pro-0813 instead? Swap to deepseek-v4-pro and price-check against the Pro rates before rollout.

The API is stateless — your client must resend the full message history with every request; only the web and app surfaces keep conversation state on DeepSeek’s side. That has not changed. Background reading: DeepSeek OpenAI SDK compatibility.

Licensing and hardware

Timeline chart of the DeepSeek V4 Preview period: the V4 Preview launch and press coverage on Apr 24, 2026, through the retirement of deepseek-chat and deepseek-reasoner on Jul 24, 2026 at 15:59 UTC

Both V4 Preview models shipped under the MIT license for weights and code — as do their successors DeepSeek-V4-Pro-0813 and DeepSeek-V4.1-Flash-1017, and earlier V3.2, V3.1 and R1. Some older releases (V3 base, Coder-V2, VL2) still split MIT code from a separate DeepSeek Model License on weights, so check the specific Hugging Face repo if licensing matters for your deployment. On the training side, CNN reports that DeepSeek partnered with Huawei on the compute side, with Huawei confirming it supports V4 via its Supernode technology combining Ascend 950 clusters; Counterpoint’s Wei Sun highlighted that V4 was run on domestic chips from Huawei and Cambricon, in contrast with R1’s Nvidia training, arguing this could give V4 an even bigger impact than R1 by accelerating domestic adoption.

What V4 does not fix

Three honest caveats from the April preview. First, the benchmark gap on expert-level reasoning was real: on DeepSeek’s preview model card HLE put V4-Pro (Preview) at 37.7% — below Claude (40.0%), GPT-5.4 (39.8%) and well below Gemini-3.1-Pro (44.4%). Second, factual world-knowledge retrieval trailed Gemini on SimpleQA-Verified by a wide margin (57.9 vs 75.6 on the same card). Third, the “Preview” label meant exactly what it said — the preview models, their pricing and their routing have all since been replaced. Keep your eval harness warm. For a broader look at where DeepSeek slots against the competition, see DeepSeek vs Claude and DeepSeek vs ChatGPT, or the full DeepSeek news hub for the next beat in the story.

What is the DeepSeek V4 release date?

DeepSeek V4 Preview released on April 24, 2026, with two open-weight model IDs going live the same day: deepseek-v4-pro (preview: 1.6T total / 49B active parameters) and deepseek-v4-flash (preview: 284B total / 13B active), both with a one-million-token context window under the MIT license. The rollout covered Hugging Face weights, the API, and chat.deepseek.com simultaneously. The official releases followed: DeepSeek-V4-Flash-0731 (API public beta, 31 July 2026), DeepSeek-V4-Pro-0813 (13 August 2026), DeepSeek-V4.1-Flash (10 September 2026) and DeepSeek-V4.1-Flash-1017 (17 October 2026), per DeepSeek’s Change Log. See the DeepSeek V4 model page for tier details.

How do I migrate from deepseek-chat or deepseek-reasoner to V4?

Keep your base_url and API key and swap the model field to deepseek-flash or deepseek-v4-pro. Thinking is the default (the old reasoner behaviour); disable it with extra_body={"thinking": {"type": "disabled"}} for the old chat behaviour. The legacy IDs were scheduled for retirement on July 24, 2026, 15:59 UTC and are no longer listed in DeepSeek’s docs. Step-by-step in our DeepSeek API best practices guide.

Is DeepSeek V4 free to use?

The web chat at chat.deepseek.com is free for individual use; DeepSeek’s 17 October 2026 Change Log says its web and app models were updated to DeepSeek-V4.1-Flash-1017. The open weights are free to download under MIT and run locally if you have the hardware. API usage is metered per token — as of October 2026, deepseek-flash costs $0.30 / $0.15 per 1M cache-miss input tokens and $1.20 / $0.60 per 1M output tokens (peak / off-peak), and deepseek-v4-pro $1.32 / $0.66 and $3.96 / $1.98. More on the free surface at is DeepSeek free.

Does DeepSeek V4 beat GPT-5.4 or Claude Opus 4.6?

On the April 2026 preview’s benchmarks, no — not across the board. On DeepSeek’s preview model card, V4-Pro (Preview, Max mode) leads Claude on Terminal-Bench 2.0 and LiveCodeBench, sits within 0.2 points of Claude and level with Gemini on SWE-Bench Verified, trails all three on HLE, and trails Claude and GPT-5.4 on HMMT 2026 math. DeepSeek-V4-Pro-0813 is reported on different benchmarks, so these numbers do not carry over. The pitch is cost, not supremacy. Full head-to-heads at DeepSeek comparisons.

Can I run DeepSeek V4 locally?

The April preview weights were about 160 GB (Flash) and 865 GB (Pro) on Hugging Face. The current releases are larger downloads: DeepSeek-V4.1-Flash-1017 about 510 GB and DeepSeek-V4-Pro-0813 about 893 GB, so both realistically need a multi-GPU server or aggressive quantization to serve. Both are published under MIT on Hugging Face. For a practical setup walkthrough, see install DeepSeek locally.

Sources & Methodology

Last source check: October 27, 2026

Primary sources

Technical references

Context sources

  • NewsBloomberg: DeepSeek unveils newest flagship a year after AI breakthroughLaunch-day coverage of V4 Flash / Pro debutLast checked: October 27, 2026
  • AnalysisSimon Willison: V4 launch day writeupIndependent confirmation of V4 sizing and pricingLast checked: October 27, 2026
  • NewsCNN coverage of DeepSeek V4 PreviewFraming V4 as preview rival to OpenAI/Anthropic/GoogleLast checked: October 27, 2026
  • NewsOfficeChai: DeepSeek V4 independent benchmark reportingCross-check on IMOAnswerBench/HMMT/SWE-Verified scoresLast checked: October 27, 2026

Methodology

Information in this article was checked against the official DeepSeek documentation, the article's cited primary sources, and the deepseekai.guide pricing dataset on the date shown above. Where third-party numbers are quoted, the originating source is linked inline so the reader can verify and re-check.

Data confidence

High for facts cited inline with primary-source links; medium for context narrative.

Editorial note

AI tooling moves quickly. If you spot an outdated number or a broken link, the contact link in the footer is the fastest way to flag it for the editorial team.

Leave a Reply

Your email address will not be published. Required fields are marked *