Seek-Chat: DeepSeek V4: Pro, Flash & Vision Exp API Pricing

DeepSeek V4 began as the two-model Preview family DeepSeek-V4-Pro and DeepSeek-V4-Flash. The current hosted documentation now lists three API IDs. deepseek-v4-flash serves the text-only DeepSeek-V4-Flash-0731 release in public beta; deepseek-v4-pro serves the larger text-only DeepSeek-V4-Pro-0813 release in GA; and deepseek-v4-flash-vision-exp serves the experimental multimodal DeepSeek-V4-Flash-Vision-Exp model announced August 21, 2026. Vision Exp accepts text with images and uses the same published token prices as V4 Flash. All three list a one-million-token context window, 384K maximum output, and Responses API support.

DeepSeek controls its official web service, apps, API, model repositories, availability, and prices.

DeepSeek V4 at a glance

Fact	Officially documented information
Release naming	Original family: DeepSeek-V4 Preview; current hosted versions: DeepSeek-V4-Flash-0731, DeepSeek-V4-Pro-0813, and experimental DeepSeek-V4-Flash-Vision-Exp
Text-model architecture	Mixture of Experts (MoE)
V4 Flash size	284B total parameters; 13B activated
V4 Pro size	1.6T total parameters; 49B activated
Context length	1M tokens for all three currently documented API models
Maximum API output	384K tokens for all three
Reasoning	Thinking and non-thinking are listed for all three; the effort mapping discussed below is documented for Flash and Pro
API formats	OpenAI-compatible Chat Completions, Responses API, and Anthropic-compatible access
Documented modality	V4 Flash and V4 Pro are text only; Vision Exp accepts text with JPEG, PNG, GIF, or WebP images
Published checkpoints	Pro, Flash, Pro Base, and Flash Base; this review does not assert a downloadable Vision Exp checkpoint
Weights license	MIT License for the published Pro and Flash checkpoints

The API model IDs and downloadable checkpoint names serve different purposes. deepseek-v4-pro, deepseek-v4-flash, and deepseek-v4-flash-vision-exp identify hosted API models. Names ending in -Base identify downloadable foundation checkpoints, not API model IDs. The experimental Vision Exp listing does not by itself establish downloadable weights. Similarly, “Max” is a reasoning-effort setting rather than a separate model endpoint.

DeepSeek V4 text models: Pro vs Flash

For text-only work, start with V4 Flash when cost, throughput, or request volume matters most; evaluate V4 Pro when harder reasoning, knowledge, coding, or agent performance can justify its higher token price and lower default concurrency. Flash is not merely a quantized copy of Pro: they are separate text models with different total and activated parameter counts. If the request includes images, use the separate experimental Vision Exp ID rather than sending image content to Flash or Pro.

Specification	DeepSeek V4 Flash	DeepSeek V4 Pro
API model ID	deepseek-v4-flash	deepseek-v4-pro
Current hosted version	DeepSeek-V4-Flash-0731	DeepSeek-V4-Pro-0813
Release status	Public beta	GA
Total parameters	284B	1.6T
Activated parameters	13B	49B
Context length	1M tokens	1M tokens
Maximum API output	384K tokens	384K tokens
Thinking and non-thinking	Supported	Supported
Responses API	Supported	Supported
Default account concurrency	2,500 active connections	500 active connections
Post-trained weight precision	Mixed FP4 and FP8	Mixed FP4 and FP8
Good starting fit	High-volume text processing, routine reasoning, and cost-sensitive applications	Hard reasoning, complex coding, knowledge-heavy work, and difficult agent tasks

DeepSeek V4 Pro: specifications and good use cases

DeepSeek V4 Pro is the larger model from the original Preview family, with 1.6 trillion total parameters and 49 billion activated for each token. On August 13, 2026, DeepSeek rolled out DeepSeek-V4-Pro-0813 as GA on App, Web, and API behind deepseek-v4-pro. DeepSeek positions it for difficult reasoning, knowledge, coding, and agent workloads. Pro is the stronger model to test first when an incorrect result is more expensive than the additional API cost, or when a task contains long chains of dependent decisions.

• Complex repository analysis, debugging, and multi-step software agents.
• Knowledge-heavy research and difficult STEM reasoning.
• Long-horizon agents that must plan, call tools, and recover from failures.
• Selected High or Max reasoning requests where evaluation shows a measurable gain.

Pro should not be selected automatically for every request. Its listed output price is higher than Flash, and its default account concurrency is lower. A production router can send routine work to Flash and escalate only tasks that benefit from Pro in repeatable tests.

DeepSeek V4 Flash: specifications and good use cases

DeepSeek V4 Flash has 284 billion total parameters and 13 billion activated per token. The hosted deepseek-v4-flash ID serves DeepSeek-V4-Flash-0731 in public beta. Flash and Pro now both support the native Responses API. Flash’s smaller active footprint, lower official API price, and higher default concurrency make it the practical starting point for routine chat, extraction, classification, summarization, coding agents, and first-pass agent work.

• High-volume classification, extraction, summarization, and support workflows.
• Cost-sensitive chat and content-processing applications.
• First-stage routing before difficult requests are escalated to Pro.
• Behavior-preserving migration from the older deepseek-chat and deepseek-reasoner aliases.

Flash can approach Pro on some reasoning tasks when given a larger thinking budget, but that does not make their behavior identical. Test both models on your documents, prompts, failure cases, latency targets, and budget before changing a production route.

DeepSeek V4 Flash Vision Exp: experimental image input

deepseek-v4-flash-vision-exp is the experimental multimodal API model DeepSeek announced on August 21, 2026. It accepts text together with JPEG, PNG, GIF, or WebP images. The two existing IDs—deepseek-v4-flash and deepseek-v4-pro—remain text only. DeepSeek says Vision Exp’s pure-text capability is on par with V4 Flash; that is a provider claim, not an independent result from this site.

Vision Exp item	Current documented value
Status	Experimental
Image delivery	Base64 data URL, external URL, or Files API file_id
Files API	DeepSeek documents upload and reuse by file_id; the Files API itself is free, while inference remains billable
Formats	JPEG, PNG, GIF, WebP
API surfaces	Chat Completions, Responses, and Anthropic-compatible messages
Modes and features	Thinking and non-thinking; JSON Output, Tool Calls, Chat Prefix Completion beta, Responses, and Anthropic-compatible access
FIM Completion	Listed as not supported. A September 7, 2026 test nonetheless returned HTTP 200 with a valid completion; treat that as undocumented behavior, not a contract
Context / maximum output	1M / 384K tokens
Listed concurrency	2,500
Image billing	Images are input tokens; up to 384 input tokens per image after documented resizing
Off-peak price / 1M	$0.007 cache hit / $0.22 cache miss / $0.66 output
Peak price / 1M	$0.014 cache hit / $0.44 cache miss / $1.32 output

Because the model is experimental, evaluate image accuracy, privacy handling, latency, tool behavior, and failure cases before production use. The complete Vision Exp guide covers request schemas, Files API use, image limits, pricing, and examples.

DeepSeek V4 specifications and API capabilities

The table below compares the two text-only official API models. Vision Exp has the separate feature summary above because its image blocks and FIM support differ. OpenRouter, cloud platforms, other hosted providers, and self-hosted runtimes can expose different model names, feature sets, context limits, and operational behavior.

Capability	V4 Flash	V4 Pro	Important condition
Thinking mode	Yes	Yes	Enabled by default
Non-thinking mode	Yes	Yes	Disable thinking explicitly
JSON Output	Yes	Yes	Chat supports json_object; Responses also supports json_schema. Validate outputs in application code.
Tool calls	Yes	Yes	Your application executes function tools; Responses also offers a server-side web_search tool
Strict tool calls	Beta	Beta	Requires the /beta base URL and supports a documented JSON Schema subset
Chat Prefix Completion	Beta	Beta	Requires the /beta URL and an assistant message with prefix: true
FIM Completion	Beta	Beta	Non-thinking only; 4K maximum output; requires the /beta URL
Context caching	Yes	Yes	Enabled by default; only qualifying repeated prefixes receive cache-hit pricing
OpenAI Chat Completions format	Yes	Yes	Change the base URL and model ID
Native Responses API	Yes	Yes	For these two text-only IDs, resend full history and do not send image input; Vision Exp supports input_image separately
Anthropic-compatible format	Yes	Yes	Compatibility does not imply support for every Anthropic field

Current FIM contract (rechecked September 7, 2026): DeepSeek’s Models & Pricing matrix lists FIM Completion (Beta) for deepseek-v4-flash and deepseek-v4-pro in non-thinking mode only, and lists deepseek-v4-flash-vision-exp as unsupported. Use the Beta base URL https://api.deepseek.com/beta with the /completions endpoint; FIM has a 4K maximum output. The specialized FIM guide uses deepseek-v4-pro in its example, and the FIM API reference enumerates only deepseek-v4-pro as a possible model value; neither overrides the current model matrix. We tested all three IDs against https://api.deepseek.com/beta/completions on September 7, 2026 and every one returned HTTP 200 with a valid completion, including Vision Exp, which the matrix marks unsupported. Flash is therefore usable for FIM. Vision Exp works too, but that is undocumented behavior which may be withdrawn without notice, so build against the documented matrix rather than our observation, and runtime-test the chosen model before production. The same position is stated on the pricing page.

One-million-token context and 384K maximum output

All three currently documented API models list a one-million-token context window and a maximum output of 384K tokens. The prompt and generated completion must fit within the context window; these figures should not be added together as a 1.384M-token total. One million tokens also does not mean one million words or guarantee perfect retrieval from a long prompt.

For independent hosted-API evidence beyond DeepSeek-published scores, see our measured DeepSeek V4 long-context results across four prompt tiers, four objective tasks, latency, token usage, and cost.

Long-context results improve when documents are filtered, labeled, ordered, and divided into meaningful sections. Retrieval, citations, and post-generation verification remain important. See the separate DeepSeek V4 context and output limits guide for token planning and failure cases.

Thinking, non-thinking, High, and Max

Thinking is enabled by default, with high as the default effort. The current Chat Completions mapping is identical for Flash and Pro: low maps to low, medium maps to high, high maps to high, xhigh maps to high, and max maps to max. The directly documented Chat values are low, high, and max; medium and xhigh are compatibility inputs. The Responses API uses reasoning.effort with none, low, high, or max; none disables thinking. Do not carry Chat compatibility inputs such as medium or xhigh, or an unsupported minimal value, into Responses requests. Recheck the official Thinking Mode guide before pinning production behavior. Max is a configuration, not a separate model endpoint.

In thinking mode, temperature and top_p have no effect. The current Chat Completions reference separately marks presence_penalty and frequency_penalty as deprecated and ineffective. When a Chat Completions request includes tools, preserve and resend the full reasoning_content from every assistant turn in every subsequent request between user messages—even when that turn did not produce a tool call—or the API returns HTTP 400. When the request does not include tools, earlier reasoning_content can be omitted. Review the official thinking-mode rules and our implementation-focused DeepSeek thinking mode guide before deploying multi-turn agents.

Default account concurrency

DeepSeek’s current model matrix lists per-account concurrency of 2,500 active connections for V4 Flash, 500 for V4 Pro, and 2,500 for Vision Exp. These are concurrent connections, not requests per minute or requests per second. All API keys under one account share the limits. Exceeding the applicable allowance returns HTTP 429. DeepSeek also documents a no-additional-cost capacity request process, subject to its assessment of business needs. See the official rate-limit and isolation documentation.

DeepSeek V4 pricing snapshot

Current price schedule verified: August 24, 2026. Prices are in U.S. dollars per one million tokens. Current off-peak and peak rows appear first; the historical pre-cutover rows apply only to Flash and Pro and are retained as dated reference.

Rate period	Official API model	Cached input	Uncached input	Output
Current off-peak	deepseek-v4-flash	$0.007	$0.22	$0.66
Current peak	deepseek-v4-flash	$0.014	$0.44	$1.32
Current off-peak	deepseek-v4-flash-vision-exp	$0.007	$0.22	$0.66
Current peak	deepseek-v4-flash-vision-exp	$0.014	$0.44	$1.32
Current off-peak	deepseek-v4-pro	$0.022	$0.66	$1.98
Current peak	deepseek-v4-pro	$0.044	$1.32	$3.96
Historical — through August 16, 2026, 15:59 UTC	deepseek-v4-flash (DeepSeek-V4-Flash-0731)	$0.0028	$0.14	$0.28
Historical — through August 16, 2026, 15:59 UTC	deepseek-v4-pro (DeepSeek-V4-Pro-0813)	$0.003625	$0.435	$0.87

Under the current schedule, peak hours are 01:00–04:00 and 06:00–10:00 UTC (09:00–12:00 and 14:00–18:00 Beijing time) from Monday through Friday only. Every other time—including all weekend hours—uses the off-peak rates. Peak rates are twice the current off-peak rates. Cache-hit pricing applies only to qualifying cached input tokens. Vision Exp images are billed as input tokens, with an upper bound of 384 input tokens per image after documented resizing. Thinking has no separate token rate, but longer reasoning can increase billed output. Check the official model and pricing table before approving a production budget, or use the DeepSeek pricing guide and calculator for per-request estimates.

How to access DeepSeek V4

Official web and app access

DeepSeek’s release notice states that V4 Preview is available through its web service, mobile apps, and API. Consumer-interface labels such as Expert Mode and Instant Mode belong to DeepSeek’s products. Do not assume that every interface label maps permanently and one-to-one to Pro or Flash unless the interface identifies the underlying model.

Official API model IDs and base URLs

• V4 Flash: deepseek-v4-flash
• V4 Pro: deepseek-v4-pro
• V4 Flash Vision Exp: deepseek-v4-flash-vision-exp — experimental image input
• OpenAI-compatible base URL: https://api.deepseek.com
• Anthropic-compatible base URL: https://api.deepseek.com/anthropic

Seek-Chat.com’s last authenticated GET /models request on July 28, 2026 identified only the Flash and Pro text endpoints; that dated response predates Vision Exp. DeepSeek’s current Models & Pricing page and August 21, 2026 changelog add the experimental Vision Exp ID, while the static List Models reference example still shows two. No authenticated August 21, 2026 rerun was made for this editorial update. Read DeepSeek model names and API IDs to distinguish product labels, hosted API names, downloadable checkpoints, and provider-specific aliases.

This minimal request uses V4 Flash with thinking explicitly disabled:

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {
        "role": "user",
        "content": "Summarize the differences between DeepSeek V4 Pro and Flash."
      }
    ],
    "thinking": {"type": "disabled"},
    "stream": false
  }'

For authentication, Node.js, Python, streaming, error handling, retries, and production safeguards, use the complete DeepSeek API documentation.

DeepSeek V4 API IDs and post-cutoff alias status

Legacy alias	Documented behavior during transition	Required explicit replacement
deepseek-chat	V4 Flash, non-thinking	deepseek-v4-flash with thinking disabled
deepseek-reasoner	V4 Flash, thinking	deepseek-v4-flash with thinking enabled

Do not migrate deepseek-reasoner automatically to Pro. Its behavior-preserving target is V4 Flash with thinking enabled. Test Pro separately where its quality advantage can justify the price and lower default concurrency. The complete DeepSeek V4 migration guide covers request changes and regression checks.

DeepSeek V4 download: official model weights

DeepSeek publishes four V4 repositories through the verified deepseek-ai organization on Hugging Face. The post-trained models use mixed FP4 and FP8 precision, while the Base checkpoints use mixed FP8 precision.

Official repository	Purpose	Parameters
DeepSeek-V4-Flash	Post-trained Flash model	284B total / 13B active
DeepSeek-V4-Pro	Post-trained Pro model	1.6T total / 49B active
DeepSeek-V4-Flash-Base	Foundation checkpoint for research and adaptation	284B total / 13B active
DeepSeek-V4-Pro-Base	Foundation checkpoint for research and adaptation	1.6T total / 49B active

“Open-weight” is the precise description for the published Pro, Flash, Pro Base, and Flash Base checkpoints: those weights are downloadable under the MIT License. The hosted Vision Exp listing does not by itself establish a downloadable checkpoint or license. Published weights do not mean every training dataset, complete training pipeline, hosted service component, or provider deployment has been released. Use the official organization to avoid mislabeled uploads, and verify the publisher, conversion method, license, and checksums of third-party derivatives.

Can you run DeepSeek V4 locally?

Yes, because official weights and inference instructions are available. However, neither full checkpoint is an ordinary laptop model. Flash has 284B total parameters, and Pro has 1.6T. Storage size is not the same as RAM or VRAM, and long contexts add substantial cache and serving requirements.

There is no single trustworthy “minimum GPU” figure for every precision, quantization, context length, batch size, runtime, and parallelism design. Self-hosted context and output limits depend on the serving stack and available infrastructure rather than automatically inheriting the official API limits. Use the local DeepSeek installation guide for deployment planning.

What changed from DeepSeek V3.2 and R1?

V4 keeps the Mixture-of-Experts approach and Multi-Token Prediction design used by earlier DeepSeek models while introducing architectural and training changes intended to make very long contexts more efficient.

Technical component	Role described in the DeepSeek report
Compressed Sparse Attention (CSA)	Compresses key-value entries and applies sparse selection to reduce long-context work
Heavily Compressed Attention (HCA)	Uses more aggressive key-value compression while retaining dense attention
Manifold-Constrained Hyper-Connections (mHC)	Strengthens residual connections while constraining signal propagation for stability
Muon optimizer	Supports faster convergence and training stability
DeepSeekMoE and on-policy distillation	Routes tokens through selected experts and consolidates domain-specialist capabilities

DeepSeek reports that Pro, at a one-million-token context, uses 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek V3.2. It reports 10% of the FLOPs and 7% of the KV cache for Flash under the same comparison. These are developer-reported efficiency figures, not measurements reproduced by Seek-Chat.com. Architecture, training, precision, and evaluation details appear in the DeepSeek V4 technical report.

V4’s hosted transition also combines general responses, thinking, long-context processing, and tool use inside the Pro and Flash family. Historical R1 and V3.2 checkpoints can still matter for reproducible research and existing self-hosted systems. See the DeepSeek models guide when choosing across model families.

DeepSeek V4 benchmarks: official and independent results

DeepSeek-published Max results

The following scores were published by DeepSeek for the Max reasoning configuration. They are not independent Seek-Chat.com measurements. Scores can change with prompts, reasoning budgets, tools, scaffolding, test harnesses, and evaluation rules.

Benchmark	V4 Flash Max	V4 Pro Max	What it tests
GPQA Diamond (Pass@1)	88.1	90.1	Graduate-level science reasoning
LiveCodeBench-v6 (Pass@1-CoT)	91.6	93.5	Competitive coding
SWE Verified (Resolved)	79.0	80.6	Repository issue resolution
Terminal Bench 2.0 (Acc)	56.9	67.9	Agent work in terminal environments
MRCR 1M (MMR)	78.7	83.5	Retrieval across a million-token context

The table supports a practical, limited conclusion: Pro shows a clearer advantage on difficult knowledge and agent tasks, while Flash can remain close on several reasoning and coding evaluations when given a larger reasoning budget. It does not establish that either model is the strongest option for every workload.

Independent evaluation: what CAISI found

In an independent evaluation, the U.S. Center for AI Standards and Innovation described DeepSeek V4 Pro as the most capable PRC model it had evaluated. CAISI found a wider gap from leading U.S. systems on several held-out and agentic tests than DeepSeek’s selected benchmark comparisons suggested. It also found V4 Pro more cost-efficient than GPT-5.4 mini on five of seven evaluated tasks, with relative cost ranging from 53% lower to 41% higher depending on the task. Source: CAISI Evaluation of DeepSeek V4 Pro, NIST, published May 1, 2026; link rechecked September 7, 2026.

Those findings should not be mixed directly with DeepSeek’s table. They used CAISI’s serving setup, system prompts, scaffolding, reasoning settings, token budgets, and prices available during its evaluation. Review the full CAISI evaluation of DeepSeek V4 Pro for its methodology and caveats.

DeepSeek V4 capabilities and limitations

Where V4 can fit well

• Agentic coding: repository analysis, debugging, planning, tool use, and multi-step software tasks.
• Long-document workflows: technical documentation, policies, logs, transcripts, research collections, and contract sets.
• Reasoning: mathematics, science, planning, and structured analysis.
• High-volume text processing: classification, extraction, summarization, and support workflows routed through Flash.
• Structured applications: JSON responses and tool requests combined with application-side validation.
• Open-weight research: experimentation or self-hosting where suitable infrastructure is available.

Important limitations

• Different release labels: DeepSeek-V4-Flash-0731 is a text-only public-beta release, DeepSeek-V4-Pro-0813 is a text-only GA release, and DeepSeek-V4-Flash-Vision-Exp is experimental. Behavior, availability, limits, and prices can still change.
• Model-specific modality: V4 Flash and V4 Pro remain text only. Vision Exp accepts documented image input but is experimental; the reviewed guide does not list video input or image generation. Do not confuse any hosted API ID with Janus, DeepSeek-VL, or OCR weights.
• Company-published benchmarks: selected public scores do not guarantee results on your workload.
• Large local footprint: official checkpoints require substantial storage, memory, and serving infrastructure.
• Long-context fallibility: a 1M window does not eliminate missed details, hallucinations, or prompt-injection risk.
• Endpoint-specific JSON contracts: Chat Completions JSON Output uses json_object; Responses also documents json_schema. Both still require application-side validation and safe failure handling.
• Tool execution: the model requests a tool call; your application validates the arguments, executes the tool, and returns the result.
• Privacy and governance: do not send secrets, credentials, regulated records, or personal data without an approved processing, access, logging, and retention design.

Review what not to paste into DeepSeek before using any hosted model with sensitive information.

Which DeepSeek V4 model should you choose?

Workload	Starting choice	Reason
Routine text chat, extraction, or classification	Flash, non-thinking	Lower cost and less unnecessary reasoning
Screenshot, chart, photograph, or document-image understanding	Vision Exp	The experimental hosted V4 model that accepts image input; evaluate it before production use
High-volume production API	Flash	Lower token prices and higher listed default concurrency
Long-document summarization	Flash first	Test Pro only where evaluation shows a quality gain
Complex coding or repository repair	Pro, High	Stronger starting position for difficult coding and agent work
Science, mathematics, or difficult planning	Pro, High	More active capacity for knowledge-heavy reasoning
Unusually hard boundary tests	Pro, Max	Uses the largest documented reasoning budget
Cost-sensitive agent workflow	Flash, High	Evaluate whether it meets the task before escalation
Local research	Flash weights first	Smaller than Pro, although still infrastructure-intensive

Route by modality first: image requests require Vision Exp, while Flash and Pro are text only. Within text workloads, a measurable routing policy is usually better than selecting Max or Pro for every request. Start with the least expensive configuration that passes your acceptance criteria, record failures, and escalate only the categories that show a repeatable benefit.

Frequently asked questions

What is DeepSeek V4?

DeepSeek V4 is a Mixture-of-Experts family introduced as the V4 Pro and V4 Flash Preview models on April 24, 2026. DeepSeek later released the text-only Flash-0731 public beta and Pro-0813 GA services, then added the experimental deepseek-v4-flash-vision-exp text-and-image API model on August 21, 2026. All three currently documented API IDs list thinking and non-thinking modes, a 1M-token context window, 384K maximum output, and Responses support. The published MIT-licensed downloadable checkpoints are the Pro and Flash lines; this page does not assert Vision Exp weights.

Is DeepSeek V4 officially released?

Yes, with different status labels. DeepSeek-V4-Flash-0731 is a text-only public-beta API release, DeepSeek-V4-Pro-0813 became text-only GA on App, Web, and API August 13, 2026, and DeepSeek-V4-Flash-Vision-Exp is an experimental API model announced August 21, 2026. The original April 24, 2026 announcement remains relevant to the Preview architecture and published Pro/Flash weights.

What is the difference between DeepSeek V4 Pro and Flash?

Pro has 1.6T total parameters and 49B activated parameters and is positioned for harder text reasoning, coding, knowledge, and agent tasks. Flash has 284B total and 13B activated and is positioned for lower-cost, higher-throughput text work. Both list a 1M context and 384K maximum output. Neither accepts native image input; use the separate experimental Vision Exp model when images are required.

What are the DeepSeek V4 API model names?

DeepSeek’s current Models & Pricing page and August 21, 2026 changelog list deepseek-v4-pro, deepseek-v4-flash, and deepseek-v4-flash-vision-exp. The first two are text only; the third is experimental and multimodal. Pro Max, Flash Max, and non-thinking are configurations rather than separate API model names.

Does DeepSeek V4 support a one-million-token context?

Yes. DeepSeek lists a 1M-token context length and 384K maximum output for Pro, Flash, and Vision Exp. The prompt and completion must fit within that context, and the size does not guarantee perfect recall of every detail.

Where can I download DeepSeek V4?

Use the official Pro, Flash, Pro Base, and Flash Base repository links in the download table above. The August 21, 2026 API listing does not by itself establish downloadable Vision Exp weights. Verify any third-party Vision Exp claim or quantization independently.

Is DeepSeek V4 open source or open weight?

The precise description is open weight for the published V4 Pro, Flash, Pro Base, and Flash Base checkpoints, which DeepSeek releases under the MIT License. The current Vision Exp API listing does not establish downloadable weights or a license. Neither statement should be expanded to imply that every dataset, service component, or infrastructure layer has been released.

Does DeepSeek V4 support images or video?

V4 Pro and V4 Flash remain text only. The separate experimental deepseek-v4-flash-vision-exp model accepts JPEG, PNG, GIF, and WebP images through Chat Completions, Responses, and Anthropic-compatible messages. The reviewed Vision guide does not list video input or image generation. See the Vision Exp image API guide.

Does DeepSeek V4 replace deepseek-chat and deepseek-reasoner?

Yes for current integration planning. DeepSeek documented both strings as temporary compatibility aliases and set July 24, 2026 as their retirement cutoff. Current Models & Pricing documentation lists the explicit Flash, Pro, and experimental Vision Exp IDs. Use the model that matches the request modality and do not keep either legacy alias in new or production code.

Is DeepSeek V4 good for coding?

DeepSeek positions V4 for coding and agent workloads, and its published evaluations include competitive coding and repository-repair tests. Pro is the stronger starting choice for difficult coding agents, while Flash can be more economical for routine generation, explanation, classification, and high-volume code tasks. Test both on your repositories before deployment.

---

Editorial verification: Current model IDs, release labels, modalities, Vision Exp features, and pricing were checked against DeepSeek’s Models & Pricing page, August 21, 2026 changelog and release note, and Vision guide. The Flash/Pro thinking-effort mapping and architecture sources retain their earlier source review; the cited independent evaluation and alias observation retain their original July 28, 2026 test date. No authenticated GET /models rerun or paid Vision Exp inference request was made for this editorial update. Hosted-provider and self-hosted behavior can differ from the official DeepSeek API.

Related DeepSeek V4 guides

• DeepSeek V4 Flash Vision Exp image API guide
• DeepSeek API documentation
• DeepSeek API pricing
• DeepSeek V4 context and output limits
• Migrate legacy aliases to DeepSeek V4
• Run DeepSeek locally
• DeepSeek model names explained