DeepSeek-V4-Pro

What it is. DeepSeek-V4-Pro is DeepSeek’s flagship open-weight model: a 1.6T-parameter Mixture-of-Experts (MoE) with 49B active parameters, released under the MIT license. It served on the API from the V4 Preview (2026-04-24) and was updated to the 0813 GA checkpoint on 2026-08-13. Why it matters. It was DeepSeek’s top-tier model for agentic coding, world knowledge and math/STEM before the V4.1 family began rolling out, and it remains one of the largest open-weight MoE models in the China AI Hub database. Key characteristics. A 1,048,576-token context window, 393,216-token maximum output, thinking (default) and non-thinking modes with reasoning_effort low/high/max, JSON output and tool calling — but no vision. What a professional should know. Deprecation was announced 2026-09-10, and the official pages disagree on whether the V4-Pro API continues unchanged after 2026-09-14 or routes requests to V4.1-Flash until V4.1-Pro launches. Anyone planning production workloads should resolve that conflict against the current change log before committing.

Architecture and parameters

DeepSeek-V4-Pro is a sparse MoE: 1.6 trillion total parameters with 49 billion active per token. That 3.1% activation ratio is what makes the model economical to serve — only a small expert subset computes on any given input, so per-token cost is driven by active, not total, parameters. The trade-off, standard for MoE designs, is that total parameters still dictate memory and self-hosting footprint, which is why a 1.6T model is a multi-GPU deployment even though inference is sparse. See the site’s Mixture-of-Experts explainer and the Chinese MoE architectures research for how this activation dial compares across labs.

What the context window actually means

A 1,048,576-token context means V4-Pro can hold roughly 700,000–800,000 English words of working material in a single request — enough for whole-repository code review, multi-document legal or financial analysis, or very long agent traces. The 393,216-token output ceiling is large but a fraction of input, so V4-Pro is best read as a long-input analysis tool rather than a long-form generation tool. See the context-window research for how this input/output split fits the wider market.

Pricing implications

List pricing is off-peak $0.66 input / $1.98 output per 1M tokens; peak windows are 2x (see the known-limitations list). The 3x input-to-output ratio is typical — output tokens bill at a premium because they consume more inference capacity. For long-context work this matters doubly: filling a 1M-token window and generating long outputs multiplies cost quickly, so cache-hit pricing and batch scheduling are where real savings live. Prices trace to the DeepSeek pricing page and were verified 2026-09-20.

API, coding, and agent implications

V4-Pro was DeepSeek’s agentic-coding flagship: vendor-reported Terminal-Bench 2.1 (87.9) and DeepSWE (62.7) place it in the top tier of the database for terminal and repository-scale agent tasks. It exposes JSON output and tool calls, so it slots directly into function-calling and agent harnesses — the same machinery the DeepSeek Harness product builds on. The deprecation notice is the operational caveat: with the V4.1 family superseding it, new integrations should prefer DeepSeek-V4.1-Flash unless a specific V4-Pro checkpoint is required. See the DeepSeek vs Kimi K3 and DeepSeek vs Qwen3.8-Max comparisons for workload-level differences.

Open weights and license

V4-Pro is MIT-licensed open weight and self-hostable. MIT is the most permissive standard license in the database — no attribution or revenue thresholds, unlike the conditional custom licenses used elsewhere. The practical caveat is verification: the Hugging Face checkpoint was last modified 2026-06-22, and it is not documented whether it matches the 0813 GA API checkpoint. Self-hosters should pin against that uncertainty. See licensing explained and open weight vs API.

Benchmark interpretation

All four listed scores are vendor-reported. HLE 42.7 (60.0 with tools), Terminal-Bench 2.1 87.9, DeepSWE 62.7, and Agents’ Last Exam 25.7 describe a strong coding/agentic profile with a modest absolute score on the hardest knowledge exam. The gap between the HLE base score and its with-tools score is itself informative: tool access materially raises the ceiling, which is the pattern behind agentic evaluation generally.

What the benchmarks do not prove

These numbers do not establish that V4-Pro beats any specific competitor, and they cannot be compared across vendors without matching benchmark versions, evaluation harnesses and tool setups. The source_type is vendor_reported for every row — none are independently re-measured. See how to read vendor-reported benchmarks and the benchmark-methodology research for the comparability traps.

Suitable and less suitable workloads

Well-suited: agentic coding, repository-scale software work, long-document analysis, math/STEM reasoning, and any open-weight deployment where a permissive license and self-hosting matter. Less suited: vision and multimodal tasks (not supported), low-latency or low-cost chat where the flash tier is cheaper, and any production dependency that cannot tolerate the post-2026-09-14 deprecation ambiguity.

Where this model fits

Workload	Relevance
Long-context analysis	High
Coding	High
Structured API workflows	High
Agent orchestration	High
Local self-hosted deployment	High
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates that V4-Pro’s situation — a large, MIT-licensed flagship marked deprecated while its API status is in flux — illustrates how quickly China’s frontier layer turns over. Its off-peak $0.66/$1.98 pricing and 49B-active MoE design made it a strong cost/performance point at launch, but the V4.1 family’s asymmetric, cheaper architecture superseded it within months. The model remains relevant to self-hosters who want a permissive 1.6T checkpoint, but API buyers should treat it as legacy and verify the current routing before reliance.

What is uncertain

• The deprecation status is conflicting: the news page says V4-Pro routes to V4.1-Flash after 2026-09-14, but the same-day change log says the API continues unchanged.
• Whether the Hugging Face checkpoint (last modified 2026-06-22) matches the 0813 GA checkpoint is not documented.
• Serving regions are not disclosed.
• All four benchmark scores are vendor-reported; no independent third-party measurement is recorded.

Sources

• DeepSeek API docs — Models & Pricing
• DeepSeek API Change Log
• DeepSeek-V4 Preview release
• Hugging Face model card — DeepSeek-V4-Pro

Labels used above: Official fact (from primary sources), Vendor-reported claim (benchmark scores published by DeepSeek), and China AI Hub analysis (our synthesis, always introduced as such). No third-party benchmark evidence is currently recorded for this model.