DeepSeek-V4-Flash

DeepSeek-V4-Flash is a relatively compact yet high-performance model in the DeepSeek-V4 family, built on a Mixture-of-Experts architecture with 284 billion total parameters, of which only 13 billion are activated per token. It supports a context length of up to one million tokens and, at the same time, requires significantly fewer resources for deployment than the Pro version. The model was trained on more than 32 trillion tokens and underwent two-stage post-training, and its MIT license makes it fully open for commercial and research use.

The key engineering innovation of V4 is a hybrid attention mechanism that radically changes the approach to long context. Unlike previous versions, where MLA was applied uniformly, in V4 different layers use two modes: Compressed Sparse Attention and Heavily Compressed Attention. CSA compresses context tokens into a compact vector at a 1:4 ratio, after which the Lightning Indexer selects only the most relevant compressed blocks for attention computation. HCA applies extreme 1:128 compression and performs full global attention over super-compact representations of the entire history, without using sparse selection. In parallel, in every layer a sliding window of 128 tokens processes the immediate local context without compression, preserving detailed local awareness. On a million-token context, this combination reduces KV-cache memory costs by nearly a factor of ten.

The model supports three reasoning modes: Non-Think for fast everyday responses, Think High for deliberate logical analysis, and Think Max for solving the most complex tasks. In terms of task-solving quality, Flash in maximum reasoning mode delivers results close to Pro-Max, despite having three times fewer activated parameters. On Codeforces the model scores 3052 points, matching Gemini-3.1-Pro-High. On LiveCodeBench it reaches 91.6% pass@1, and on Apex Shortlist — 85.7%, confirming strong abilities in programming and mathematics. Even in the basic no‑thinking mode the model holds a high bar: MMLU-Pro stands at 83.0%, reflecting a solid foundation of general world knowledge.

The model is suited for scenarios where both context length and deployment cost-efficiency are critical. These include analysis and summarization of large documents, legal dossiers, or scientific articles; search across extensive corporate knowledge bases; multi-step agent work with long sessions and many tools; and programming and mathematical modelling tasks where both reasoning accuracy and the ability to keep voluminous code or specifications in view are important. Developers can use it as a powerful and resource-accessible alternative to closed flagship models for building chatbots, automatic code generation systems, intelligent document assistants, and research agents that operate on contexts of impressive size.

Announce Date: 22.04.2026
Parameters: 293B
Experts: 256
Activated at inference: 13B
Context: 1049K
Layers: 43
Attention Type: DeepSeek Sparse Attention
Developer: DeepSeek
Transformers Version: 4.57.1
License: MIT

Use our pre-built public endpoints for free to test inference and explore DeepSeek-V4-Flash capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.