Author: Peng Cheng
An LLM all-in-one appliance is an integrated device that combines AI acceleration hardware (GPUs / domestic AI accelerator cards), server hardware, inference frameworks, large model weights, and supporting O&M and application platforms. It achieves out-of-the-box usability, enabling governments, enterprises, and research institutions to deploy localized private large models, thereby eliminating the complex processes of users assembling servers, adapting frameworks, and debugging models on their own.
Early last year, DeepSeek R1 was officially unveiled, driving the rapid popularity of DeepSeek all-in-one appliances. The DeepSeek all-in-one appliance can deploy the full-scale 671B large model in a single-machine environment. For a time, institutions, enterprises, public institutions, and SMEs across the country rushed to purchase them, leading to high market enthusiasm.
Today, however, driven by market shifts and model evolution, the LLM all-in-one appliance market is undergoing profound changes.
01. The Highly Controversial DeepSeek All-in-One Appliance
Around the Spring Festival in 2025, DeepSeek R1 emerged, and its related popularity continued to climb. Subsequently, over 60 manufacturers launched DeepSeek all-in-one appliances. However, this was followed by various debates, with everyone discussing whether relevant units should deploy these appliances. After all, a single appliance can cost over a million CNY, resulting in huge financial expenditures. In contrast, calling cloud-based large models would cost so little that the budget could last for centuries. Moreover, as large models continue to iterate and parameter counts keep expanding, the computing power limits of single-machine hardware will gradually become apparent.
In just three months, the market cooled down. An industry consensus has formed: DeepSeek is useful, but the DeepSeek all-in-one appliance is truly useless. After all, a standard DeepSeek all-in-one appliance requires at least 8 GPU cards. Limited by the single-machine architecture, it cannot enjoy the economies of scale brought by cluster deployment, keeping the comprehensive usage costs high. More critically, many early appliances were essentially just hardware bundled products with insufficient software optimization, merely completing simple model installation and deployment. Although they claimed to support the operation of the 671B large model, their actual concurrency capability was very poor, and simultaneous access by multiple users easily caused lag.
Some manufacturers focus on hardware delivery while neglecting subsequent O&M. After equipment delivery, they lack continuous model update and iteration services, requiring users to invest additional manpower for secondary modifications. This has led to many institutions having the hardware in place but failing to run their businesses. Moreover, full-scale appliances generally cost between CNY 1.2 million and CNY 1.5 million. When using domestic cards, since the VRAM of a single unit is insufficient, two units often need to be connected in parallel, bringing the budget to nearly CNY 4 million. By the second half of 2025, "full-scale versions are hard to sell" was no longer a secret, and the main order-winning products for small and medium-sized manufacturers quietly shifted to 32B and 70B models.
Moving into 2026, the market environment has changed once again. Memory, storage (SSDs, HDDs, etc.), GPU chips, and other components have seen successive price increases. From an asset perspective, appliances purchased in early 2025 can now be resold in the second-hand market to recover the entire procurement cost and even generate a profit. A netizen revealed that their company sold five sets of Ascend card servers to the R&D department of a certain state-owned enterprise last year, with a transaction amount of just over CNY 10 million. This year, upon learning from the client's project manager that the project failed and the project team was disbanded, they achieved profitability by selling the five servers to an adjacent project team. The project team even received a bonus, becoming a rare team that made a profit despite project failure. The purchasers, who were initially mocked by netizens for falling into the trap of high prices, have instead become beneficiaries in this round of hardware price increase cycle.
However, the hardware appreciation phenomenon needs to be viewed objectively. Although the hardware has appreciated on paper, it does not mean it is suitable for procurement in current new projects. Second-hand appliances generally face practical issues such as expired warranties, firmware and driver compatibility, and excessively high power consumption. To run new-generation MoE models like DeepSeek V4.1 Flash on older hardware architectures, inference framework modifications are still required. It is not a case of directly running the full-scale model right after receiving the equipment, which also inhibits institutions' willingness to hoard second-hand equipment on a large scale.
02. LLM All-in-One Appliances Collectively "Scale Down"
The industry trend has shifted this year. The market no longer chases appliances capable of running the full-scale 671B large model; instead, lightweight local deployment devices like AI BOXes are emerging. Compared to the high-end appliances for the full-scale 671B model last year, the hardware configurations and the supported model parameter counts of appliances on the market this year have been significantly reduced, with most mid-range devices capable of running at most 70B-level models.
The reasons for this scale-down are not complicated. First, the demand side is squeezing out the bubble. A large amount of practice has proven that the vast majority of users need scenarios such as knowledge base Q&A and official document writing. For these scenarios, distilled small models are sufficient, and the capabilities of the671B model are mostly overqualified. Second, the capability density on the model side is improving. Technologies such as MoE sparse activation, distillation, and quantification have matured, significantly reducing the number of parameters required to achieve the same effect. Third, the supply of high-end chips is restricted. In April 2025, the H20 chip was indefinitely suspended from sale. At that time, domestic accelerator cards had improved in VRAM and throughput performance through iteration, but a single card still could not run ultra-large MoE models. To achieve full-scale operation, multiple devices need to be connected in parallel. In addition to procurement costs, multi-card parallel connection brings issues such as complex debugging and a heavy workload for software ecosystem adaptation, further raising the threshold for project implementation. Fourth, memory prices have risen, making the cost of locally deploying the full-scale version too high and lacking cost-effectiveness.
Consequently, the entire all-in-one appliance market has been split in two: the lightweight, scaled-down versions focus on high volume to support the overall market shipments; full-scale high-end appliance products still exist, but the purchasing groups are highly concentrated in central and state-owned enterprises and financial institutions with strict compliance requirements, leading to a significant decline in the overall shipments of high-end models.
03. Rising Memory and Token Prices, Surging Popularity of Agents
At a time when all-in-one appliances are generally moving towards lightweight designs, do high-end appliances still have market value? In fact, high-end appliances still have market demand in specific scenarios.
First, in complex production scenarios, small models have clear shortcomings. Agent long-chain tasks, Vibe Coding engineering-level code development, and AI professional video generation—these three fastest-growing AI applications currently—have extremely high requirements for the model's reasoning depth, logical chains, and multimodal understanding capabilities. To stably support such production-level tasks, full-scale large models must be deployed. It is precisely for this reason that high-end all-in-one appliances have not disappeared due to the popularity of lightweight products; instead, their value is even more pronounced in actual production processes.
Second, cloud API calling costs are soaring, highlighting the cost-effectiveness of local deployment. As Token prices rise, coupled with the massive consumption of multimodal generation, the cost of using cloud-based large models on a large scale is climbing rapidly. Vibe Coding requires repeated iterations and massive generation for debugging; AI video generation is even billed by the second, with the cost of a single video ranging from several to dozens of CNY. When the business volume increases, cloud subscription fees will be astonishing, and annual investments can easily exceed the million-CNY level, making the cost advantage of locally deployed large models re-emerge. At the same time, the importance of data security is becoming increasingly prominent. Just recently, security incidents involving the exposure of user data were reported overseas, and overseas closed-source AI products have begun to embed mechanisms such as invisible watermarks. Enterprises and institutions continue to strengthen their demands for data localization.
Finally, the popularization of Agents boosts the utilization rate of local deployments. In the past, the most criticized aspect of all-in-one appliances was insufficient utilization, with million-CNY-level equipment sitting idle most of the time. However, the popularity of Agents allows all-in-one appliances to work continuously, causing the utilization rate to surge, which theoretically can increase equipment utilization. Through deep integration with business systems via workflow orchestration, all-in-one appliances can host continuously running Agent tasks, achieving 24/7 uninterrupted inference.
04. Open-Source Model Updates Bring a Turning Point for Appliance Capability Upgrades
The true industry turning point comes from the iteration of new-generation models, as the capabilities of large models themselves continue to strengthen. Although Kimi K3 has outstanding performance, its massive parameter count makes it difficult to deploy on all-in-one appliance hardware. The emergence of new models has changed this situation.
First, DeepSeek V4.1 Flash was officially released, offering Flash-tier usage costs with capabilities close to the Pro version. According to official introductions, V4.1 Flash comprehensively surpasses V4 Pro in multiple metrics such as performance, overhead, inference speed, and total task time. The model adopts a brand-new architecture, natively supports visual understanding, and takes over the positioning of the previous generation Pro model. Crucially, V4.1 Flash is built on the CausalEncoderDecoder (Causal Encoder-Decoder) architecture and is an MoE (Mixture of Experts) model with a total parameter count of 552B.
It should be noted that the 552B model size is smaller than the 671B parameter scale of the full-scale DeepSeek R1 model. This means that the full-scale appliance purchased for CNY 1.5 million last year can be upgraded in place without adding a single card, with its capabilities enhanced in place. The capabilities are stronger, the context reaches the latest level, and the model has not even become larger. Moreover, the model is jointly trained with the Harness framework, enhancing its long-cycle continuous task processing capabilities.
Second, the MiniMax H3 open-source model quickly gained popularity. MiniMax H3 is a general-purpose all-modality generation model released by MiniMax, officially launched on July 31, 2026. On August 3, 2026, MiniMax officially open-sourced the model. Its open-source license explicitly excludes usage rights in the United States, the European Union, the United Kingdom, and South Korea. Domestic and international chip manufacturers such as Huawei Ascend, Moore Threads, MetaX, Hygon, Kunlunxin, Iluvatar CoreX, Biren Technology, AMD, and Intel, as well as development communities and cloud inference platforms like Hugging Face, ModelScope, ComfyUI, RunningHub, and fal, along with inference frameworks such as vLLM-Omni and SGLang, have completed adaptation and support. The model ranks first in video editing capabilities on the Artificial Analysis leaderboard.
There is no need to elaborate on how expensive AI video generation from cloud service providers is. The outstanding comprehensive capabilities of the H3 model make local generation of AI videos a reality, reducing enterprises' reliance on commercial cloud video generation services. Moreover, after community optimization, MiniMax H3 can even run on a configuration of 9700x + 32GB memory + 5070ti, generating a 5-second video in about 5 minutes. It is conceivable that after optimization, the output speed of all-in-one appliances will be even faster. A recent message reported that for generating a 5-second, 1344×768 video on 4 RTX 6000D cards, the original BF16 50-step scheme took 348.8 seconds, nearly 6 minutes. After full acceleration by RunningHub H3 Lightning, it only takes 28.7 seconds. This represents an approximately 12-fold speedup, reducing generation time by about 92%, while still retaining BF16 numerical precision. RunningHub has open-sourced the entire workflow on GitHub, making it out-of-the-box usable for any creator! In an environment with 8 RTX 6000D cards, having H3 generate a 15-second video with a resolution of 768×1344 takes about 48 seconds for pure text generation and about 73 seconds for dual-reference image generation. A 15-second image-to-video can now hit the time scale of "results within 1 minute."
05. Conclusion
In summary, purchasers who acquired high-end all-in-one appliances in the first half of last year have gained asset-level returns in this round of hardware price increases. In 2026, the market shows polarization: the hardware costs of full-scale high-end all-in-one appliances have risen, and their cost-effectiveness has declined. Procurement is only recommended for institutions with rigid business and compliance isolation needs. At the same time, it must be noted that purchasing all-in-one appliances is not a one-time fix. In addition to the hardware procurement budget, institutions need to continuously bear hidden costs such as data center power supply and cooling, O&M manpower, model version iteration, inference framework tuning, and fault maintenance. Many units only calculate hardware expenditures and ignore subsequent O&M expenses, ultimately causing project implementation to fall short of expectations. Lightweight AI BOX devices are gaining popularity and are more suitable for business scenarios with local confidentiality demands and low AI capability requirements.
In the future, the all-in-one appliance industry will not simply move towards "getting smaller and smaller"; the product tiering landscape will continue to evolve. On the one hand, lightweight AI BOXes will further penetrate the market, covering more edge and departmental-level businesses. On the other hand, as domestic chips and MoE inference frameworks continue to iterate and mature, the comprehensive costs of high-end full-scale all-in-one appliances will have room to drop in the future. Going forward, competition in the all-in-one appliance track will no longer be about who can deploy models with the largest parameter counts, but rather the comprehensive strength of the entire solution encompassing hardware, inference frameworks, model adaptation, business applications, and O&M services.