DeepSeek V4.5

In-depth exploration of the iterative version of the flagship large model launched in July 2026, strengthening code, Chinese long text and multi-round consistency, continuing the open source and floor-price play style

large language model LLM DeepSeek

https://www.deepseek.com/

Visit Website

In-depth Report

DeepSeek V4.5 is an iterative version of the flagship large model launched by DeepSeek in early July 2026. Its positioning continues the consistent strategy of the V4 family of "open source, ultra-long context, and floor price". This version mainly focuses on three directions: code generation, Chinese long text understanding and multi-round dialogue consistency. It is still open to the web and app for free, and the API side reuses the call entrance of the V4 series. It is not a re-telling of the big story, but more like a pragmatic update that "pushes the already strong areas to a higher level." It is especially friendly to cost-sensitive developers and heavy text workers.
The company behind DeepSeek is DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. located in Hangzhou. It has taken the open source route since 2023 and has successively released DeepSeek LLM, Coder, Math, V2, V3, R1 and other series, relying on extreme training and inference efficiency to gain a foothold in the open source camp. On April 24, 2026, the V4 preview version was officially launched and open sourced simultaneously. Two specifications were given in one breath: V4-Pro (approximately 1.6 trillion total parameters, approximately 49 billion per activation) and the cost-effective V4-Flash (approximately 284 billion total parameters, approximately 13 billion per activation), which benchmarks top closed source, both are fully open sourced under the MIT protocol. V4.5 was released in a low-key manner within the next two to three months. The community weekly report showed that its launch time would be around July 2, and there was no press conference, continuing DeepSeek's style of "no notice, just launch".
V4.5 follows the V4 family's three-level reasoning intensity design: non-think mode (Non-think), regular deep thinking (Think High), and the strongest thinking (Think Max). Developers can adjust the intensity through the reasoning_effort parameter of the API. For complex Agent scenarios, the official recommendation is to turn on Max directly. At the architectural level, the V4 system introduces a hybrid attention mechanism that compresses the token dimension, combined with the self-developed DSA sparse attention, Engram conditional memory (which separates static knowledge from dynamic reasoning, and the long text retrieval accuracy increases from 84.2% to 97.0%), and MHC manifold constrained hyperconnection (which amplifies inter-layer signals with trillions of parameters to within 2 times), thus significantly reducing computing power and graphics memory overhead in long contexts. The context window of the V4 series reaches 1 million tokens and is set as standard for the entire series; while the aggregation model catalog records V4.5 as 256K context and a maximum of 33K output. The specific caliber is subject to official follow-up announcements. According to community feedback, V4.5’s code generation (HumanEval was mentioned to reach a new high), Chinese long text understanding and multi-round consistency are the three most obvious improvements this time.
DeepSeek’s pricing philosophy has always been to “shoot the bottom line.” The web chat.deepseek.com and the official app provide free conversations for individual users. This is the fundamental weapon that dislocates it from the paid subscriptions of ChatGPT and Claude. The API side is billed by token. The official pricing page shows that V4-Flash costs 0.02 yuan per million token input (cache hits), 1 yuan for misses, and 2 yuan for output; V4-Pro costs 0.025 yuan, 3 yuan, and 6 yuan respectively, and the concurrency limit is 2500/500. The overseas aggregation directory's hosting API quotation for V4.5 is US$0.5 for input and US$1.1 for output per million tokens, which is still the floor price in RMB. One time point to note is that the old deepseek-chat and deepseek-reasoner model names will be deprecated on July 24, 2026, pointing to the non-thinking and thinking modes of V4-Flash respectively. The production environment needs to complete the model parameter migration before then. Model weights are open sourced in HuggingFace and ModelScope under the MIT license and can be deployed locally or in private clouds, leaving only infrastructure as the marginal cost.
In the real experience of ordinary users, the positive parts are quite consistent. 1 Million Contexts has been repeatedly praised as "too practical" - throwing a dozen documents into it at once and letting it read through and then answer, the labor cost of long contracts, financial reports, and entire book analysis has been reduced from thousands to dozens of yuan. The coding ability has also been recognized. The Flash version is sufficient for daily scripts, and the Pro version is more stable in handling complex tasks. Some users reported that a batch order processing script ran through in 15 minutes. Cheapness is another trump card. The monthly bill for intensive use is less than 30 yuan, while similar foreign models often start at 200 to 300 yuan. The Chinese language sense is also considered to be better than that of foreign models, and it can accurately understand local expressions such as "Maulika must die". Negative feedback was equally concentrated. The most commonly named one is the "weather card problem": when it is asked to check structured data (weather, stock prices), it will occasionally compile numbers. It is fine when it is an assistant, but it will overturn when it is an authority. There are still gaps in complex common sense reasoning. The classic "Desperate Father" red-green color blindness genetic problem V4 failed to capture the core logic in the first round. The effective utilization of longer contexts has also been questioned. According to third-party testing, it can only stably identify tags within about 100,000 tokens in a million-word text, and the effective content is far less than a quarter of what is claimed. In addition, there are "overthinking" and redundant answers on simple tasks, and the essence of the model is still pure text without multi-modal capabilities.
Looking at the global camp, DeepSeek V4 won the best open source model in the Agentic Coding evaluation. The actual test experience of internal employees was better than Sonnet 4.5, and the delivery quality was close to the Opus 4.6 non-thinking mode, but there is still a gap with the Opus 4.6 thinking mode; the world knowledge test was significantly ahead of the open source model, only slightly inferior to Gemini-Pro-3.1; the mathematics and competition codes were comparable to the top closed source. Its real power lies in cost: it reduces the price of GPT-5.5 from about US$30 per million output to less than US$1, which is called a "head-on collision of pricing philosophies" in the industry. More importantly, V4 has made full-link adaptation for domestic computing power (Huawei Ascend 950PR, Cambrian) from the bottom architecture. The inference speed of domestic chips is about 35 times higher than that of V3. It is regarded as a key link in China's AI full-stack independent controllable closed loop.
There are several controversial points worth noting around V4.5. The first is the gap in effective context: 1 million tokens are claimed, but the actual usability is far less than this, which may not be as good as expected for users who rely on the "whole library code, whole book" scenario. The second is the lack of multi-modality: while competing products in the same period regard vision and video as their core selling points, DeepSeek is still a pure text model and lags behind in complex multi-modal Agent scenarios. Third, there are few public benchmarks for V4.5 - the specific running scores that can be found at present are mostly from aggregate directories and community weekly reports. Authoritative third-party evaluations are not sufficient, and they need to be marked as pending verification when citing. The fourth is security performance: In the jailbreak test, it can complete basic interception, but it lacks the security reasoning awareness to actively identify hijacking instructions. The wording is "cannot" instead of "won't", and there is still room for polishing the credibility details. The fifth is the geopolitical dimension: Affected by U.S. export controls, pre-training still relies on NVIDIA, but post-training and inference are deeply tied to domestic chips. This is not only a strategic asset, but also means that production capacity is strongly related to the pace of price cuts and the supply of domestic computing power.
If you are a heavy text worker - editor, researcher, student, lawyer, analyst, long document processing capabilities can save a lot of time; if you are a developer, the efficiency and accuracy of code assistance, debugging, writing comments, and generating test cases are all in the first echelon; if you are a cost-sensitive individual user, the free version is completely sufficient. Scenarios that are not recommended are: Those who require advanced image/video generation or complex multi-modal tasks, currently DeepSeek cannot satisfy them; those who pursue the most advanced in-depth reasoning or require a mature plug-in ecosystem and enterprise-level SLA, closed-source models are still a more stable choice; those who treat AI as an authoritative data source and adopt it directly without verification, be wary of its hallucination tendency.
DeepSeek V4.5 does not tell a grand new story, but pushes the three already strong fulcrums of "easy to use, cheap, and open source" one step further, especially in terms of code, Chinese long text, and multi-round consistency, giving users a real sense of improvement. Its shortcomings are also clear: effective long context and plain text form are still the gaps between it and the top closed source. For the vast majority of Chinese users, it is a flagship model worthy of first choice and with almost no barriers to entry.

User Reviews

JerryTurnerSr59

—

The local deployment party is ecstatic, the MIT protocol directly pulls the weight, the consumer-grade graphics card is enough to run the quantitative version, and there is no need to worry about privatized deployment. Our company cannot use public cloud APIs due to data compliance requirements. We have been stuck on model authorization before. Once the V4 series MIT protocol was opened, the legal department gave the green light. Now the internal inference cluster runs the quantitative version Pro. The delay and cost are much more cost-effective than purchasing closed source APIs, and it is too friendly to company compliance and cost.

4pa0w

—

The overall flaws are not hidden, it is easy to use, cheap and open source. The domestic model can achieve this level, which can no longer be explained by the patriotic filter. I am willing to use it as the main force in daily production.

ticklishpeacock996

—

Linked with Claude Code to work as an Agent, it is good for simple tasks. In complex projects, it is occasionally disobedient and executed without confirmation. Engineering details still need to be polished. Remember to add guardrails in the production environment.

qvhqs2

—

The cost of reasoning is ridiculously controlled. The same quality of work can easily cost tens of dollars in closed source. It can be converted into RMB or less than one dollar, and the pricing is to squeeze the opponent to death. We have done some calculations and found that the customer service intention classification and summary task for the same month costs more than 2,000 US dollars using GPT, but less than 20 yuan using DeepSeek API. Small and medium-sized teams directly benefit, and the budget saved is enough to support half of the operation. Of course, whether complex reasoning should be closed source or closed source cannot be generalized.

于鹏敏

—

The only thing that gives me a headache is the "weather card problem". When I ask it to check the prices of competing products, it will occasionally make up numbers that look very reasonable. However, after checking, they are all fake. I still have to go through the key data myself.

MatthewFisher

—

The pure text model is a bit regretful. Now that everyone else is working on multi-modality, it can’t even see pictures. You still need other tools to do visual-related work. I look forward to making it up later.

Amanda.Gomez_Max3

—

There are still gaps in complex reasoning. The genetic problem of red-green color blindness did not turn around in the first round. The officials themselves admitted it. It is okay to be an assistant, not an authority. Make the key conclusions yourself.

ScottTaylor_20225

—

Let’s use the free version first. It’s great for long document analysis. I can throw more than 30 pages of PDF directly into it and ask three questions. It can really quote paragraph by paragraph, which saves a lot of money compared to foreign models.

SSimmons21

—

Millions of contexts are highly publicized, but if you actually put a whole book into it, it will still be a bit confusing later on, and the effective length is far less ideal than one million. I have done the "needle in a haystack" test myself, burying markers in a novel that is close to a million words. It can only reliably recognize the first 100,000 words or so, and the error rate becomes higher the further it goes. This point needs to be realistic. Don't be led by parameters. If you really want to deal with very long documents, you should segment them or manage the search yourself.

heavytiger177

—

The understanding of long Chinese texts in V4.5 has been significantly improved this time, and multiple rounds of conversations are not easy to go astray. The daily writing and editing experience is smoother than the previous version. I specifically use the same 5,000-word manuscript for A/B. In the previous version, I often forgot the tone I set before in the second round. This version can basically remember the detailed requirements of "keep it colloquial and don't add emojis." It is a real improvement for content workers. There is no need to re-explain the background every round, which saves a lot of prompt words.

PGonzalez

—

Writing Python data processing scripts in V4.5 is really smooth. My batch order cleaning job used to take two hours with human help, but now it can be done in fifteen minutes.

ChainMax_n

—

The coding ability is really strong, Flash is enough for daily use, Pro handles complex tasks more stably, and the team has switched to it for daily coding. Last week, I asked it to reconstruct a 3,000-line historical baggage module. It first drew the calling relationship and then changed it file by file. It was more stable than the new colleagues. The only drawback was that it occasionally wrote more defensive code, which seemed verbose, but the correctness cannot be said. There is nothing to pick in terms of cost-effectiveness. For the same quality, the price quoted by closed sources can be dozens of times higher.

Jacqueline.Edwards_2020

—

The API price is too ridiculous. Flash output is only two yuan per million tokens. I used it intensively for a month and the bill was less than thirty yuan. In the past, I used ChatGPT at least two to three hundred yuan.

Dennis_Evans766

—

It will think too much about simple questions and write a lot of verbose words. Only by turning off the thinking mode can it be more efficient. I hope that in the future, it will be smarter to judge whether to think deeply.

Jacqueline_Perez

—

The Chinese language sense is indeed better than that of GPT. When I said, "This product is dead," it immediately understood that the profit margin was low and would not make me a joke that the product was dead.

DianeGutierrez

—

It takes half an hour to generate a first draft of industry analysis with citations. It is very friendly to research. Although it is not deep enough, it can be used as a scaffolding, saving me at least two hours. I usually let it outline the framework and find public data sources first, and then do a cross-validation of internal data myself. This way I don't have to worry about it making up stuff, and I can outsource the mechanical material collection. To put it bluntly, it is a tireless junior analyst, and you can make your own conclusions.

DeepSeek V4.5

In-depth Report

User Reviews

Similar Products