On April 23, 2026, DeepSeek officially released the DeepSeek-V4 Preview and simultaneously open-sourced the model weights. Following the V3 series, this marks another major iteration on the MoE (Mixture-of-Experts) architecture. DeepSeek positions it as “entering the era of million-token context for everyone,” with 1M token context length becoming standard across all official services.

This article is based on official release notes, API documentation, the HuggingFace technical report, and initial discussions from Chinese tech communities like V2EX, summarizing V4’s core technical features, product strategy, and early market feedback.

1. Dual-Model Strategy: Pro vs. Flash

DeepSeek-V4 adopts a parallel dual-version strategy to cover different performance and cost requirements:

The Pro version targets scenarios requiring the strongest reasoning and knowledge capabilities, while Flash focuses on cost-effectiveness and fast response times. According to the official WeChat article, Pro’s Agentic Coding ability has surpassed Sonnet 4.5, with delivery quality approaching Opus 4.6 non-thinking mode, though a gap remains with Opus 4.6’s thinking mode.

2. Three Architectural Innovations

According to the HuggingFace technical report, the V4 series introduces the following key architectural upgrades:

2.1 Hybrid Attention Mechanism (CSA + HCA)

V4 designs a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), achieving significant efficiency gains in 1M token long-context scenarios. The technical report shows that DeepSeek-V4-Pro’s single-token inference FLOPs in this scenario are only 27% of V3.2’s, with KV cache reduced to just 10%.

2.2 Manifold-Constrained Hyper-Connections (mHC)

mHC reinforces traditional residual connections, enhancing cross-layer signal propagation stability while maintaining the model’s expressive power.

2.3 Muon Optimizer

The training phase uses the Muon optimizer, which the official report claims outperforms traditional approaches in both convergence speed and training stability.

3. Training and Post-Training Paradigm

The V4 series was pretrained on over 32T tokens. Post-training follows a two-stage paradigm:

1. Independent domain expert cultivation: Through SFT and GRPO (Group Relative Policy Optimization) reinforcement learning, expert capabilities in different domains are trained separately
2. Unified model consolidation: Domain expert capabilities are distilled and integrated into a single model through on-policy distillation

This “train separately, then consolidate” strategy aims to bring the final model to a high level simultaneously across code, reasoning, knowledge, and other domains.

4. API and Product Changes

Model Name Updates

For API calls, the base_url remains unchanged (https://api.deepseek.com), but the model parameter needs to be updated to:

• deepseek-v4-pro
• deepseek-v4-flash

The legacy model names deepseek-chat and deepseek-reasoner will be deprecated on July 24, 2026. For compatibility, they currently map to Flash’s non-thinking and thinking modes respectively.

Thinking Mode

Both V4 versions support thinking mode, enabled by default. The API controls it through the following parameters:

• thinking: {type: "enabled/disabled"} — thinking mode toggle
• reasoning_effort: "high/max" — thinking intensity control

Thinking content is returned via the reasoning_content field, at the same level as content. Note that thinking mode does not support temperature, top_p, presence_penalty, or frequency_penalty parameters.

For detailed documentation, refer to the DeepSeek Thinking Mode Guide.

5. Initial Community Feedback

After V4’s release, the V2EX deepseek node and HuggingFace discussion forums quickly filled with discussions. Here are the main feedback categories as of noon on April 24:

Positive Feedback

• Long-awaited arrival: “Rumors since last December, finally officially released”
• Architecture recognition: The 1.6T parameter scale combined with MoE architecture is considered “worth exploring its logical potential”
• Open-source attitude: HuggingFace discussions saw waves of supportive comments from users
• Long-context accessibility: 1M context becoming standard is seen as a positive signal for lowering the barrier to long-text applications

Concerns and Skepticism

• Pricing controversy: Pro version pricing (input ¥12/1M tokens, output ¥24/1M tokens) is considered high by some users; “only Flash is affordable” became a common refrain
• Gap with closed-source models: The team itself acknowledges a gap with Opus 4.6 thinking mode; some users bluntly stated “GPT-5.5 hit it right out of the gate”
• Web version experience: Some users pointed out that the official Web interface actually uses Flash for regular users, and the “expert mode” is not Pro either; the early web experience “didn’t meet expectations in absolute performance”
• Practical value questioned: “Pro seems quite expensive but not that much better” — the intuitive impression of some early adopters

Technical Discussions

• Local deployment hardware requirements became a hot topic, with FP8/FP4 quantization schemes sparking discussion
• Inference framework compatibility: the community is watching when vLLM, SGLang, and other frameworks will support V4
• Ascend support: some users mentioned “Ascend is supported,” but specific adaptation details remain unclear

6. Ascend Support Status

Regarding Huawei Ascend chip adaptation, current information is limited:

• Official stance: DeepSeek has not provided official support or adaptation plans for Ascend chips in its public repositories. In related GitHub issues, official contributors have responded to users with “please seek more support from Huawei Ascend,” suggesting that current adaptation is primarily driven by third parties or the community
• Community developments: V2EX discussions include mentions of V4 “supporting Ascend,” but lack specific technical details (such as supported Ascend models, inference frameworks, performance data, etc.)
• Media coverage: As of publication, no major Chinese tech media outlets have published in-depth reports or performance benchmarks on DeepSeek V4’s official Ascend adaptation

If your business depends on Ascend deployment, it’s recommended to monitor the Huawei Ascend official community (MindSpore or Ascend Developer Community) for relevant inference adaptation solutions, rather than waiting for DeepSeek to provide native support.

7. Summary

DeepSeek-V4’s release continues the team’s core strategy of “open-source + high cost-effectiveness.” The democratization of 1M context, MoE architecture efficiency optimization, and the Pro/Flash dual-version coverage form the fundamentals of its product competitiveness.

However, based on early community feedback, the following points are worth monitoring:

1. Pricing sensitivity: Pro version pricing has increased significantly compared to the V3 era. Whether users will pay the premium depends on whether its actual performance can demonstrate a significant difference from Flash in complex tasks
2. Competition with closed-source models: The team acknowledges a gap with Opus 4.6 thinking mode, while closed-source models like GPT-5.5 continue to iterate at an accelerating pace
3. Ecosystem adaptation: Ascend support, inference framework integration, and other ecosystem building is still in its early stages, which may affect enterprise adoption decisions

In-depth benchmark comparisons (with GPT-5.x, Claude Opus 4.x, GLM-5.1, etc.) are expected to emerge within the next 1-3 days, at which point a more accurate assessment of V4’s true competitiveness can be made.

References

• DeepSeek Official WeChat Release
• DeepSeek API Documentation
• DeepSeek API Pricing
• DeepSeek Thinking Mode Guide
• HuggingFace DeepSeek-V4 Model Page
• DeepSeek-V4 Technical Report PDF
• V2EX DeepSeek V4 Discussion Thread