On April 24, 2026, DeepSeek released DeepSeek V4, a pair of open-weight mixture-of-experts (MoE) models with a 1-million-token context window: V4-Pro, with 1.6 trillion total parameters, and the smaller V4-Flash. The benchmark scores are competitive rather than record-setting, but the architecture changes how much a long context costs, which matters most for AI agents that accumulate tool results over hours of work. As with DeepSeek-R1, the weights are published under the permissive MIT license.

Key Facts#

  • Release: April 24, 2026, with four checkpoints on Hugging Face: V4-Pro, V4-Flash and a base model for each.
  • DeepSeek-V4-Pro: 1.6T total parameters with 49B active per token.
  • DeepSeek-V4-Flash: 284B total parameters with 13B active per token.
  • Context window: 1M tokens for both models.
  • Efficiency: at 1M tokens, V4-Pro needs 27 percent of the per-token inference FLOPs and 10 percent of the KV cache memory of DeepSeek-V3.2. V4-Flash needs 10 percent of the FLOPs and 7 percent of the KV cache.
  • Precision: the instruct models store MoE expert weights in FP4 and everything else in FP8. The base models are FP8 throughout.
  • License: MIT, consistent with DeepSeek's earlier releases. Hugging Face's summer 2026 report notes that DeepSeek ships models up to 1.65T parameters under plain MIT terms.

What Happened#

DeepSeek published the models together with a technical report, and Hugging Face published a same-day analysis titled "a million-token context that agents can actually use." Its central point is that a 1M-token window is only capacity. Whether you can afford to use it depends on the cost of every forward pass at that depth, and agents pay that cost repeatedly because every tool result is appended to the context.

V4 attacks that cost with a hybrid attention design. Compressed Sparse Attention (CSA) pools every four tokens into one compressed key-value entry, then a lightweight "lightning indexer" running in FP4 picks the most relevant compressed blocks for each query. It extends the DeepSeek Sparse Attention idea that DeepSeek introduced in V3.2. Heavily Compressed Attention (HCA) compresses by a factor of 128 and attends densely to the short result. Layers alternate between the two, and both keep a sliding-window branch for the most recent uncompressed tokens. Most KV entries are stored in FP8. According to Hugging Face, the result is a KV cache roughly 2 percent the size of a conventional grouped-query attention model with eight KV heads stored in BF16.

The release also targets agent behavior directly. V4 keeps its reasoning across user turns when a conversation contains tool calls, so a multi-turn agent no longer has to rebuild its plan after every follow-up message. It introduces a dedicated |DSML| token and an XML-based tool-call format that separates string parameters from structured JSON parameters, which avoids common escaping and type errors. The instruct models offer three reasoning modes: Non-think, Think High and Think Max. Think Max needs a context window of at least 384K tokens, and DeepSeek recommends a temperature of 1.0 and a top-p of 1.0 in all modes.

On benchmarks, DeepSeek's reported results for V4-Pro in Think Max mode put it close to frontier proprietary models on agent tasks. It scored 80.6 on SWE-bench Verified, against 80.8 for Claude Opus 4.6 in its maximum mode, and 73.6 on MCPAtlas Public. On Terminal Bench 2.0, it scored 67.9, behind GPT-5.4 at its xHigh setting (75.1). These are vendor-reported numbers from the technical report.

Background#

DeepSeek drew worldwide attention with DeepSeek-R1 in January 2025, an MIT-licensed reasoning model trained at comparatively low cost. It then moved steadily toward efficiency at long context. DeepSeek-V3.2-Exp, released in September 2025, introduced DeepSeek Sparse Attention, and DeepSeek-V3.2 followed. The company also open-sourced much of its infrastructure, including the FlashMLA attention kernels, the DeepEP communication library and the 3FS distributed file system, and its V4 report describes DeepSeek Elastic Compute (DSec), an internal platform that runs hundreds of thousands of concurrent sandboxes for agent training.

V4 arrived in a crowded 2026 field. Its own comparison tables include Moonshot's Kimi K2.6 and Z.ai's GLM-5.1, Alibaba's Qwen family covered every size from phones to servers, and Google's Gemma 4 targeted the small end. DeepSeek's bet was not the highest benchmark score but the cheapest usable long context.

Why It Matters for Developers#

Long-running agents become cheaper to operate. If you build coding agents, research agents or operations bots with Microsoft Agent Framework or Semantic Kernel, context growth is often the real cost driver. A model that needs a tenth of the KV cache at depth changes how many concurrent sessions a GPU can hold, and how long an agent can run before you have to summarize or truncate its history. Measure it: log tokens per session and cost per completed task, as described in our guide to LLM observability and cost control.

Do not treat 1M tokens as a replacement for retrieval, though. Hugging Face reports that V4-Pro's 8-needle retrieval accuracy stays above 0.82 up to 256K tokens but falls to 0.59 at 1M. For large knowledge bases, a RAG pipeline remains more reliable, and the long window is best used for the working state of a task.

Tool-calling integration needs attention. V4's reasoning retention across turns only helps if your client sends the reasoning content back with the conversation history, so check how your SDK and middleware handle reasoning messages. The new XML-based tool format means self-hosted stacks need a serving runtime whose parser understands it. When you use V4 through an OpenAI-compatible endpoint from .NET, keep the model name, endpoint and sampling settings in configuration, and run a function-calling regression suite before switching production traffic.

Hardware still limits self-hosting. V4-Pro at 1.6T parameters is a multi-GPU cluster deployment. V4-Flash is more approachable: Hugging Face's summer report notes that community GGUF builds of the roughly 284B model already run in llama.cpp-based setups. For most .NET teams, the practical path is a hosted API or a managed platform first, with self-hosting reserved for data that cannot leave your infrastructure.

What's Next#

The open question Hugging Face raised at launch is how quickly the tool ecosystem adapts to V4's |DSML| schema, and whether its interleaved-thinking gains carry over to third-party agent frameworks. DeepSeek has since pushed further into agent tooling. In August 2026, it open-sourced DeepSeek Harness, a plugin-based agent harness that Ollama lists among its supported integrations, and it has continued to publish kernel libraries for sparse attention.

For architects, V4 is a reminder to design agents around context economics. Our guide to AI agent architecture patterns covers memory, summarization and tool design choices that keep agents efficient on any model.

Sources#