If you were waiting for a firm DeepSeek V4 release date, here it is: 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. After months of rumour (Hunter Alpha on OpenRouter, Reuters leaks about Huawei Ascend hardware, an “any day now” consensus on X), DeepSeek published the 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 actually is, how pricing lands against the previous generation, the benchmark picture in context, and a concrete migration plan off the legacy `deepseek-chat` and `deepseek-reasoner` IDs before they retire on July 24, 2026.

The confirmed DeepSeek V4 release date

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 model IDs: DeepSeek-V4-Pro at 1.6T total / 49B active parameters, and DeepSeek-V4-Flash 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 is not cosmetic. DeepSeek has historically used it to signal that pricing, routing, and tokenizer details can still shift during the window, so treat the numbers below as current as of the release day and re-check before you commit a production budget.

What V4 actually is: two tiers, one feature set

V4 is not a single model — it is a family of two open-weight MoE checkpoints that share architecture, context length, and the same three reasoning-effort modes. You pick the tier with the model field and the effort with a parameter.

Simon Willison’s day-of writeup independently confirmed the 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 V4-Pro the largest open-weight model currently 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 Hugging Face model card describes a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA); in the 1M-token setting, DeepSeek-V4-Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache compared with V3.2. V4-Flash drops these further: roughly 10% of the FLOPs and 7% of the KV cache of V3.2.

Pricing — current rates and the off-peak status

Pricing landed well below the previous generation. Simon Willison’s writeup quotes DeepSeek’s pricing page at $0.14/M input and $0.28/M output for Flash, and $1.74/M input and $3.48/M output for Pro. Cache-hit rates carry over the structure from V3.2 — $0.028/M for Flash and $0.145/M for Pro on the hit tier. All numbers in USD per million tokens, verified as of April 24, 2026; the V3.2 era’s off-peak discount ended on September 5, 2025 and was not reintroduced with V4.

Pro sits at roughly 6× Flash on output, so Flash remains the default recommendation for most chat and standard coding work. Pro earns its keep on agentic trajectories, long tool-use chains, and the harder SWE-Bench / Terminal-Bench style workloads where the benchmark lift justifies the spend. 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 × $0.028/M = $56.00
• Uncached input: 200 × 1,000,000 = 200M tokens × $0.14/M = $28.00
• Output: 300 × 1,000,000 = 300M tokens × $0.28/M = $84.00
• Total: $168.00

The same workload at Pro rates runs $290 + $348 + $1,044 = $1,682.00. 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

DeepSeek was clear-eyed in its own announcement. According to the V4 technical report, V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks, while falling marginally short of GPT-5.4 and Gemini-3.1-Pro — a trajectory that trails the current frontier by roughly 3 to 6 months.

Independent reporting from OfficeChai confirms the 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 — chat.deepseek.com exposes V4 behind Expert Mode (Pro) and Instant Mode (Flash). The DeepThink toggle now switches the active V4 model between non-thinking and thinking mode.
2. API — POST /chat/completions against https://api.deepseek.com, the OpenAI-compatible endpoint. An Anthropic-compatible surface is live on the same base URL.
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:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_DEEPSEEK_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)

When thinking is enabled the response returns reasoning_content alongside the final content. Core knobs worth knowing: temperature (DeepSeek recommends 0.0 for code, 1.3 for general chat, 1.5 for creative writing), 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), 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 own model card recommends setting the context window to at least 384K tokens for the Think Max reasoning mode. If you set reasoning_effort="max" without raising max_model_len, expect truncation.

Migration off legacy model IDs

If your integration still points at deepseek-chat or deepseek-reasoner, you have a hard deadline. The announcement states that deepseek-chat and deepseek-reasoner will be fully retired and inaccessible after July 24, 2026, 15:59 UTC, and currently route to deepseek-v4-flash in non-thinking and thinking modes respectively. DeepSeek’s guidance is to keep base_url and just update the model to deepseek-v4-pro or deepseek-v4-flash.

• deepseek-chat → deepseek-v4-flash (no extra parameters needed for non-thinking).
• deepseek-reasoner → deepseek-v4-flash with reasoning_effort="high" and the thinking flag in extra_body.
• Need frontier-tier agentic performance? Swap to deepseek-v4-pro and price-check against the Pro tier 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

Both V4 Preview models ship under the MIT license for weights and code — the same stance as 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. First, the benchmark gap on expert-level reasoning is real: HLE puts V4-Pro at 37.7% — below Claude (40.0%), GPT-5.4 (39.8%) and well below Gemini-3.1-Pro (44.4%), with DeepSeek acknowledging the gap directly. Second, factual world-knowledge retrieval still trails Gemini on SimpleQA-Verified by a wide margin. Third, the “Preview” label means exactly what it says — pricing, routing, and tokenizer details can still move. 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.

Last verified: 2026-04-24. DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.