DeepSeek V4 Engineering: 1M Context, MoE, and the API Migration

DeepSeek V4 is easy to describe as “the 1M-token model.” That is true, but it is not the whole story. Based on the published materials, V4 is better understood as a combined model-and-service release: two model tiers with published total and active parameter counts , a million-token context window , MoE descriptions in ecosystem materials , and API access compatible with OpenAI and Anthropic-style integrations .

DeepSeek’s Transparency Center lists V4.0 DeepSeek-V4 with a release date of April 24, 2026, and links to a Model Card and Technical Report; DeepSeek’s announcement says the DeepSeek-V4 Preview is live and open-sourced.

The confirmed lineup: Pro for headroom, Flash for efficiency

Item	DeepSeek-V4-Pro	DeepSeek-V4-Flash
Published scale	1.6T total parameters / 49B active parameters	284B total parameters / 13B active parameters
Context window	Up to 1M tokens	Up to 1M tokens
Product positioning	Largest model in the V4 family	Designed for higher-speed, higher-efficiency workloads
API model name	deepseek-v4-pro	deepseek-v4-flash

DeepSeek’s models and pricing page also lists a maximum output length of 384K for both models, along with support for features such as JSON Output and Tool Calls. That matters because V4 is not just a research-spec announcement. The Pro and Flash tiers are presented as callable products with different cost/performance expectations.

Why MoE matters: capacity and per-request compute are no longer the same thing

APIyi and HyperAI both describe V4-Pro and V4-Flash as Mixture-of-Experts, or MoE, models. In MoE systems, the total parameter count is best read as the size of the expert pool, while the active parameter count refers to the subset involved in a given inference pass. That explains why V4’s public specifications emphasize both total parameters and active parameters.

The appeal is straightforward: MoE can partially decouple model capacity from the compute used on each request. The catch is that the serving system has to handle expert routing, expert parallelism, communication, and load balancing. After the launch, the SGLang and Miles team said it had provided inference and RL training support for DeepSeek-V4, with adaptations for V4’s hybrid sparse-attention, mHC, and FP4 expert weights — a sign that the engineering challenge extends into the serving and training stack, not just the model file.

A 1M-token context window is mostly a systems problem

NVIDIA’s developer material frames V4-Pro and V4-Flash as models aimed at efficient million-token context inference, with use cases including long-context coding, document analysis, retrieval, and agentic AI workflows. DeepSeek’s API documentation also lists a 1M context length for both models.

For builders, the upside is obvious: fewer document chunks, less stitching, and potentially fewer retrieval misses. But a larger window also magnifies the pressure on attention computation, context caching, memory bandwidth, and throughput scheduling. In practice, V4 should be evaluated with real repositories, long documents, RAG pipelines, and agent toolchains — not just with a synthetic “can it accept a huge prompt?” test.

Attention terminology: the direction is clear, the labels need care

Public materials are not perfectly consistent in how they name the long-context machinery. APIyi says V4’s 1M context is powered by Hybrid Attention plus DSA sparse attention. HyperAI describes a hybrid attention design combining Compressed Sparse Attention, or CSA, and Heavily Compressed Attention, or HCA, and also mentions mHC. SGLang/Miles says its open-source stack was adapted for hybrid sparse-attention, mHC, and FP4 expert weights.

The safest reading is that the V4 ecosystem points toward some mix of sparse, compressed, and hybrid attention plus serving-stack optimization. The exact module boundaries, names, and performance contribution should be checked against DeepSeek’s official Model Card and Technical Report rather than inferred from secondary summaries alone.

API migration: the product design lowers switching friction

DeepSeek’s changelog says the API now supports V4-Pro and V4-Flash through both the OpenAI ChatCompletions interface and the Anthropic interface. To access the new models, the base_url remains unchanged and the model parameter should be set to deepseek-v4-pro or deepseek-v4-flash. DeepSeek’s first-call documentation lists the base URLs as https://api.deepseek.com for the OpenAI format and https://api.deepseek.com/anthropic for the Anthropic format.

model: deepseek-v4-pro
model: deepseek-v4-flash

OpenAI-format base_url: https://api.deepseek.com
Anthropic-format base_url: https://api.deepseek.com/anthropic

There is also a deadline for older names. The legacy API model names deepseek-chat and deepseek-reasoner are scheduled to be discontinued on July 24, 2026; during the transition period, they point to the non-thinking and thinking modes of deepseek-v4-flash, respectively.

For existing applications, the practical checklist is simple: replace the model name, choose Pro or Flash intentionally, and regression-test long context, Tool Calls, maximum output behavior, and cost under your own workload.

Claims to test, not just repeat

First, performance claims deserve caution. DeepSeek’s Chinese release page says V4-Pro reaches a leading level in China and among open-source models for agent capability, world knowledge, and reasoning; APIyi also lists benchmark claims including SWE-Verified scores. Those are useful launch claims, but they should be treated as publisher and ecosystem evidence until tested against your own prompts, budget limits, and production tasks.

Second, mechanism-level details should be separated by source quality. Terms such as Hybrid Attention, DSA, CSA, HCA, mHC, and FP4 expert weights appear across different public materials, but the naming and source hierarchy are not uniform. Outside the official technical report, it is risky to treat every label as a fully verified implementation fact.

Third, a 1M-token context window is a capacity specification, not a guarantee that every full-window request will be low-latency or low-cost. Real behavior will depend on document length, cache hits, concurrency, tool-call chains, and the evaluation standard you care about.

Bottom line

The engineering story of DeepSeek V4 is the combination: a 1.6T/49B-active Pro model, a 284B/13B-active Flash model, up to a 1M-token context window, and OpenAI/Anthropic-compatible API access in one product line. For developers, the near-term move is not to repeat the launch pitch, but to run end-to-end tests on your own long documents, codebases, RAG flows, and agent workflows — and to complete the legacy model-name migration before July 24, 2026.