If you came looking for signal rather than hot takes, the past week of DeepSeek company news actually matters for anyone running the API in production. On April 24, 2026, the Hangzhou lab shipped **DeepSeek V4 Preview** as two open-weight Mixture-of-Experts models, confirmed a July retirement date for its legacy model IDs, and — separately — entered its first external funding talks at a reported $10 billion valuation. Huawei and Cambricon silicon are now part of the training story, not just a rumour. This article pulls the verified facts together: what launched, what’s still a preview, who is buying in, and what integrators should change before July 24, 2026.

The headline: V4 Preview ships, V3.x starts its retirement clock

The biggest item in the DeepSeek company news cycle is the V4 Preview release on April 24, 2026 — DeepSeek’s first major new model in over a year. DeepSeek announced that DeepSeek-V4 Preview is officially live and open-sourced, pitching the release as the era of cost-effective 1M context length. Two model IDs shipped on day one, and they are not interchangeable marketing tiers — they are different weights with different parameter budgets:

• deepseek-v4-pro — 1.6T total / 49B active parameters, with performance rivalling the world’s top closed-source models.
• deepseek-v4-flash — 284B total / 13B active parameters, positioned as the fast, efficient, and economical choice.

Both models use the standard MIT license, which makes DeepSeek-V4-Pro arguably the largest open-weights model released to date. For production teams, the practical takeaway is the migration instruction DeepSeek published alongside the release: keep base_url, just update model to deepseek-v4-pro or deepseek-v4-flash. No endpoint rewrite, no SDK swap.

The legacy IDs are on a clock. DeepSeek stated that deepseek-chat and deepseek-reasoner will be fully retired and inaccessible after July 24, 2026 at 15:59 UTC, currently routing to deepseek-v4-flash in non-thinking and thinking mode respectively. If your integration still points at those IDs, the switch to deepseek-v4-flash or deepseek-v4-pro is a one-line change — but it is a change that must happen before that cutoff.

What V4 actually changes under the hood

V4 is not a V3.2 refresh. DeepSeek singled out a technique it calls Hybrid Attention Architecture, which it says improves the ability of the model to remember queries across long conversations, and pushed the 1 million-token context window that allows entire codebases or long documents to be sent as a single prompt. The efficiency story is where the engineering matters most — at 1M-token contexts, V4-Pro uses roughly 27% of the single-token FLOPs and 10% of the KV cache of V3.2, and V4-Flash lands at about 10% / 7% respectively.

On the API, both V4 models accept thinking mode as a request parameter (not a separate model ID). Pass reasoning_effort="high" with extra_body={"thinking": {"type": "enabled"}} to enable the thinking path; omit both to stay in the non-thinking default. Thinking requests return reasoning_content alongside the final content. Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint, and the API supports both OpenAI ChatCompletions and Anthropic APIs. Here is the minimum viable Python call against V4-Pro using the OpenAI SDK pattern:

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": "user", "content": "Summarise the V4 release notes."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
    max_tokens=4096,
)
print(resp.choices[0].message.content)

The API is stateless — clients must resend the full conversation history on every request, which is the opposite of how the web chat and mobile app behave. If you want a deeper walkthrough of the parameters, the DeepSeek API documentation covers the request shape and the DeepSeek context caching guide explains how cache-hit pricing applies.

Pricing: V4-Flash undercuts V3.2, V4-Pro is a step up

DeepSeek listed V4 rates on its public pricing page at launch. Figures below are USD per 1M tokens as of April 2026; check the DeepSeek API pricing page before committing to a spend plan, because Preview pricing can change.

DeepSeek is charging $0.14 per million tokens input and $0.28 per million tokens output for Flash, and $1.74 input / $3.48 output for Pro. Off-peak discounts ended on September 5, 2025 and were not reintroduced with V4. The historical night-rate pattern people quote from the V3 era does not apply any more.

A worked example using V4-Flash: 1,000,000 API calls with a 2,000-token cached system prompt, a 200-token uncached user message, and a 300-token response costs $56.00 (cached input) + $28.00 (uncached input) + $84.00 (output) = $168.00 total. The same workload on V4-Pro works out to roughly $1,682.00 — about 10× more, which is why Flash remains the default recommendation for general chat workloads and Pro is reserved for frontier-tier agentic or competitive-coding work. For arbitrary sizing, the DeepSeek pricing calculator runs both tiers side by side.

Funding talks: $300M at $10B, with Tencent and Alibaba in the room

Until this month, DeepSeek had taken no outside money. That changed in mid-April. In April 2026, investors began speaking with DeepSeek for a $300 million funding round, which would bring DeepSeek to a total valuation of $10 billion. On release day itself, Bloomberg reported that the two largest Chinese tech conglomerates were involved. Tencent Holdings and Alibaba Group are in talks to participate in the first funding round for DeepSeek, according to a Bloomberg report published on April 24, 2026.

The terms are not settled. Tencent has proposed acquiring up to a 20% stake in DeepSeek as part of the financing round, though the startup is reluctant to relinquish that level of control. The valuation benchmark is based on publicly traded rivals such as MiniMax, which is valued at around $40 billion. Why now? The honest answer is talent and capacity, not cash shortage. DeepSeek’s most urgent problem is talent — since the second half of 2025, several core members have left, including Wang Bingxuan (to Tencent), Wei Haoran (DeepSeek-OCR lead author), Guo Daya (DeepSeek-R1 lead author, to ByteDance), Ruan Chong (Janus-Pro, to DeepRoute.ai) and Luo Fuli (to Xiaomi).

Context on the parent company: DeepSeek founder Liang Wenfeng’s quantitative hedge fund generated returns of more than 50% in 2025, boosting the potential war chest for a company that has already shaken up the global tech landscape despite spending far less than rivals. So the $300M round is less about survival and more about pricing the company and, per industry reports, funding a commercial ecosystem that a research-only lab cannot sustain long-term. For the longer arc, see our coverage of DeepSeek funding news and DeepSeek history.

The silicon story: Huawei and Cambricon, not just Nvidia

The other thread running through this week’s announcements is hardware. To fulfil V4’s computing needs, DeepSeek partnered with Chinese tech giant Huawei, which said in a statement that it supports the AI startup with its “Supernode” technology by combining large clusters of its “Ascend 950” chips. Counterpoint Research’s Wei Sun highlighted that V4 is run on domestic chips from Huawei and Cambricon, in comparison to R1, which was trained on Nvidia hardware.

That is not a clean handover — DeepSeek said it had used both Nvidia and Huawei’s domestically-produced Ascend chips for its new model. The detail matters because US export controls shaped how the training stack was built. Reports, including in The Information, say DeepSeek has been skirting a US ban on the export of top-end chips to China; The Information reported in December that DeepSeek developed V4 using thousands of chips dismantled in third countries and smuggled to China. Nvidia told The Information it had not seen any evidence of this and that “such smuggling seems farfetched”. For the regulatory backdrop, see our tracker on DeepSeek US restrictions.

Timeline: the last six months of DeepSeek company news

Another hurdle for DeepSeek is service stability — at the end of March 2026, DeepSeek suffered an 11-hour outage that even trended on social media. Reliability, not raw model quality, is the dimension on which this next phase of the company will be judged.

Independent takeaways from running V4 in production

A few notes from actually pointing a workload at V4 since preview day, rather than reading the marketing:

• The Flash tier is the default. On general chat, RAG and tool-calling workloads, V4-Flash in non-thinking mode is the cost-efficient pick. V4-Pro earns its 10× price multiplier on agentic coding and long-horizon reasoning, not on “answer my email” traffic.
• Thinking-max needs headroom. reasoning_effort="max" wants max_model_len >= 393216 to avoid truncation. If you see empty content with a populated reasoning_content, raise max_tokens.
• JSON mode is designed, not guaranteed. Include the word “json” and an example schema in the prompt, set max_tokens generously, and handle the occasional empty response. See the DeepSeek API JSON mode reference.
• Open-weights discipline matters. Both V4 models are open source, published under the DeepSeek collection on Hugging Face; developers can download the weights, run them locally, and fine-tune them. For self-hosted deployment, see install DeepSeek locally.
• Benchmark humility. DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks, but falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails the current frontier by roughly 3 to 6 months.

Where V4 sits against the field

Any “DeepSeek vs” conversation now needs to specify the V4 tier. Flash competes on price with small frontier models; Pro competes on capability with the top tier of closed models — at far lower token cost but with a documented quality gap on the newest releases. DeepSeek called V4 the most powerful open-source platform in a challenge to rivals from OpenAI to Anthropic, unveiling V4 Flash and V4 Pro with top-tier performance in coding benchmarks and advancements in reasoning and agentic tasks. Take that framing as one side of the story; the detailed numbers belong in DeepSeek benchmarks 2026.

For head-to-heads, start with DeepSeek vs ChatGPT and DeepSeek vs Claude. If you are considering leaving the DeepSeek stack altogether, the open-source AI like DeepSeek guide sets out the realistic substitutes — Kimi K2.6 and GLM 5.1 being the most serious at the open-weights frontier.

For a fuller running log of release notes beyond the items above, see DeepSeek latest updates, or the DeepSeek news category hub for everything filed under this beat.

Bottom line

The interesting part of this news cycle is not “DeepSeek released a new model”. It is that three things moved in the same week: a model family that materially changes the context-length and price ceiling for open weights, a first external funding round at a $10B valuation benchmarked against MiniMax, and an explicit shift toward Huawei and Cambricon silicon for training. Any one of those would have been a headline. Together they describe a company transitioning out of “research lab with a viral moment” and into a commercial AI vendor with the ordinary obligations — uptime, capital structure, migration timelines — that implies.

If you are operating on the API, two concrete actions: pin your integration to deepseek-v4-flash or deepseek-v4-pro before July 24, 2026, and re-run your cost model at the new cache-hit / cache-miss / output rates for the tier you actually use. Primary sources worth bookmarking: the official V4 Preview release notes and the Bloomberg V4 coverage.

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.