DeepSeek-V4-Flash: 284B MoE for High-Throughput Pipelines at $0.14/M

DeepSeek released V4-Flash alongside V4-Pro on April 24, 2026 — not as a stripped-down version, but as a purpose-designed speed tier sharing V4-Pro's key architectural innovations. Where V4-Pro (1.6T/49B active) targets maximum quality, V4-Flash (284B/13B active) targets maximum throughput at minimum cost while retaining the two features that matter most: 1M context and the CSA+HCA hybrid attention that makes it practical.

At $0.14/M input — one-third of V4-Pro's $0.435/M and 21× cheaper than GPT-5.5 — V4-Flash targets the class of applications that generate high token volumes: document processing pipelines, RAG retrieval reranking, batch summarization, classification, and synthetic data generation. For these use cases, the speed and cost advantages outweigh the quality gap to V4-Pro.

What's New

284B/13B active MoE: V4-Flash uses the same Mixture-of-Experts design as V4-Pro at smaller scale — 284B total parameters with 13B active per token. The 13B active parameter count is roughly equivalent to a dense 13B model for inference cost purposes, but benefits from the full 284B parameter training capacity for knowledge and capability.

CSA+HCA hybrid attention carried forward: V4-Flash retains V4-Pro's core architectural innovation — Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) interleaved across layers. This gives V4-Flash the same fundamental property as V4-Pro: 1M context support with dramatically reduced KV cache and FLOPs vs a pure transformer. The efficiency gains that make 1M context practical in V4-Pro apply equally to V4-Flash.

1M context window at $0.14/M: V4-Flash supports 1M token context — the same as V4-Pro. At $0.14/M, processing a 100K token document costs $0.014. For document-heavy pipelines (contracts, research papers, codebases), V4-Flash makes full-document processing economically trivial.

MIT license: Same as V4-Pro — commercial use, fine-tuning, on-premises deployment, redistribution. V4-Flash can be self-hosted for zero marginal cost per token after infrastructure setup.

OpenAI-compatible API: deepseek-v4-flash model ID on the DeepSeek API, same format as V4-Pro and V3. Applications can route between V4-Flash and V4-Pro based on task complexity with a single model ID change.

Architecture

Shared hybrid attention architecture: V4-Flash and V4-Pro share the same CSA (Compressed Sparse Attention) + HCA (Heavily Compressed Attention) design. Both interleave CSA layers (processing standard contexts via structured sparsity) with HCA layers (achieving 90% KV cache compression for 1M context). At V4-Flash's 13B active parameters, this architecture provides long-context efficiency at speeds appropriate for real-time applications.

13B active for latency-sensitive applications: With 13B active parameters per forward pass, V4-Flash generates tokens significantly faster than V4-Pro's 49B active. For streaming applications (chatbots, voice interfaces, real-time coding suggestions) where first-token latency and tokens-per-second matter to users, V4-Flash's active parameter count is the relevant spec.

Cache economics at scale: V4-Flash cache hit price is $0.00116/M — a 120× discount from $0.14/M cache miss. For pipelines that repeatedly process the same system prompt or reference documents (RAG with static retrieval, document QA with the same source), cache hit pricing makes the effective cost far below the nominal $0.14/M.

Benchmarks

V4-Flash vs V4-Pro and comparable models:

V4-Flash's benchmarks reflect its 13B active parameter scale: competitive with 72B-class models (Qwen3-72B on SWE-bench, 52.1% vs ~52%), but below the frontier MoE tier of V4-Pro. For routing decisions: use V4-Flash for classification, summarization, extraction, and standard generation; use V4-Pro for complex reasoning, agentic coding, and tasks where quality differences are user-visible.

Pricing in Context

V4-Flash at $0.14/M with 1M context is the most capable model in the speed/cost tier at April 2026. The closest competition (GPT-4o-mini at $0.15/M) offers 128K context vs V4-Flash's 1M. Claude Haiku 4 is more expensive with shorter context. Gemini 1.5 Flash is cheaper but with lower benchmark performance.

What It Means for Developers

High-volume document pipelines: At $0.14/M with 1M context, processing 10,000 documents of 50K tokens each costs $70 total (cache miss) or $0.58 (cache hit for repeated system prompts). For batch document processing — legal review, research synthesis, content moderation — V4-Flash makes full-document LLM processing economically viable at any volume.

RAG with large retrieved contexts: V4-Flash's 1M context enables retrieval patterns that were previously impractical: retrieve 200–500 document chunks, concatenate into a single large context, and run one inference call. Compared to multiple smaller context calls with V3.2 or GPT-4o-mini, this approach simplifies architecture and can improve quality by eliminating context fragmentation.

Real-time applications with latency requirements: 13B active parameters and CSA+HCA efficiency give V4-Flash fast token generation for streaming applications. For chatbots, voice interfaces, and coding autocomplete where per-token generation speed matters, V4-Flash provides good quality at speeds appropriate for live user interactions.

V4-Flash / V4-Pro tiered routing: The same CSA+HCA architecture means both models handle long context the same way — routing between them based on task complexity is predictable. Simple extraction and summarization → V4-Flash. Complex reasoning and agentic coding → V4-Pro. Applications can implement cost-aware routing without architectural surprises.

Availability

• API: DeepSeek Platform — deepseek-v4-flash
• HuggingFace: deepseek-ai/DeepSeek-V4-Flash
• License: MIT
• Pricing: $0.14/M input, $0.28/M output, $0.00116/M cache hit

Bottom Line

V4-Flash is not a downgraded V4-Pro — it's a purpose-built speed tier sharing V4-Pro's key innovations: CSA+HCA hybrid attention and 1M context. At $0.14/M with 1M context, it's the best option in the cost-optimized tier for document-heavy applications at April 2026. For teams building V4-Pro applications, V4-Flash provides a natural cost tier for high-volume classification, summarization, and extraction tasks — same API, same architecture, one-third the price.

Resources

• HuggingFace — DeepSeek-V4-Flash
• DeepSeek API Docs — V4 API reference
• DeepSeek Platform — API access

DeepSeek-V4-Flash is available via the DeepSeek API and on HuggingFace under MIT license.