Independent notes
Free chat, a developer API, and open weights now sit on DeepSeek-V4-Pro and DeepSeek-V4-Flash, both with a 1M-token window. This is a field guide from someone who walked the official English homepage, the API docs, and Harness — not a login page.
What I actually ran
DeepSeek is a language model family. I did not find image generation or video generation on the official site, so there is no fake gallery here. What follows are notes and calls I can stand behind: chat modes, long context, Thinking, experimental vision-as-input, the API, and Harness.
Instant versus Expert on the same refactor
1M context on a long RFC
Thinking Mode on 9.11 vs 9.8
Vision-exp reading an ops chart
# Official-compatible Chat Completions call I use for V4-Pro
# base_url stays https://api.deepseek.com; only the model name changes.
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{
"model": "deepseek-v4-pro",
"messages": [
{"role": "system", "content": "You are a careful coding assistant."},
{"role": "user", "content": "Find the race in this handler."}
],
"thinking": {"type": "enabled"},
"reasoning_effort": "high",
"stream": false
}'V4-Pro Chat Completions call
from openai import OpenAI
# Same SDK I already had for OpenAI-shaped APIs.
client = OpenAI(
api_key="...", # DeepSeek API key
base_url="https://api.deepseek.com",
)
response = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[{"role": "user", "content": "Outline a migration plan."}],
reasoning_effort="high",
extra_body={"thinking": {"type": "enabled"}},
)
print(response.choices[0].message.content)OpenAI SDK pointed at api.deepseek.com
# DeepSeek Harness, developer preview — local web UI
# From the official Harness page: capabilities are plugins.
npx @deepseek-ai/dsh web
# Full source:
# git clone https://github.com/deepseek-ai/deepseek-harness
# Modes I actually saw listed: Standard, Code, Minimal, Creator.
# I treat this as a local agent runtime, not a hosted chat tab.DeepSeek Harness quick start
Primary buttons on this page open a local next-step dialog. The product itself lives on DeepSeek’s own domains.
Try nowHow I start
I open chat.deepseek.com, pick Instant or Expert, and stay in the free tab when the job is a draft or a question. Details: DeepSeek Web.
I keep base_url at api.deepseek.com and set model to deepseek-v4-pro or deepseek-v4-flash. Same SDK shape I already use. API Platform.
When I want a local agent loop, I use the developer-preview Harness — plugins for tools, sandboxes, and sessions. DeepSeek Harness.
What it is
The English homepage title is “Into the Unknown.” Under that sit two honest doors: Chat Now (free access on the web) and Access API (build with the latest models). A banner the day I visited said the official DeepSeek-V4-Pro release is live across web, app, and API, with stronger agent skills, Responses API, and Codex integration. That is the product I am writing about.
DeepSeek-V4-Pro is the heavy model: 1.6 trillion total parameters, 49 billion active, aimed at agentic coding, world knowledge, and math/STEM. DeepSeek-V4-Flash is the fast cousin: 284 billion total, 13 billion active, close reasoning, cheaper and quicker on the API. Both default to a 1M-token context on official services. Chat exposes them as Expert Mode and Instant Mode. The API uses the names deepseek-v4-pro and deepseek-v4-flash. Older names deepseek-chat and deepseek-reasoner were scheduled to retire on 24 July 2026 and, during the overlap, routed to Flash non-thinking and thinking.
I also opened the research footer: DeepSeek V4, V3.2, V3.1, R1, and V3 still have public notes. They are earlier chapters, not a second vendor. V3.2 added Thinking inside tool-use. V3.1 introduced hybrid Think / Non-Think. R1 was the MIT-licensed reasoning splash. V3 was the 671B MoE that made the free app feel fast. I keep them in the lineup so you can see the path, and I do not pretend a retired API name is a current default.
If you want the operator’s own words, use the official DeepSeek website (www.deepseek.com). This site does not host accounts, keys, or downloads. I write as a reviewer who walked through chat, docs, and Harness and then sat down to explain the doors.
Jobs I actually give it
Free web and app access for writing, Q&A, and long threads.
V4-Pro in Claude Code, OpenCode, Codex, or DeepSeek Harness.
1M context is the default on current official DeepSeek services.
Experimental vision model describes screenshots, charts, and photos you send.
Thinking Mode for decimals, proofs, and contest-style reasoning.
Skim the job, then open the product on DeepSeek’s own site. These notes stay here.
Open Browse modelsLineup
Names below match the spelling on DeepSeek’s pages and API docs. I did not translate them, and I did not add a rival’s model to fill a grid.
Flagship V4 model: 1.6T total / 49B active, 1M context, Thinking and non-Thinking, built for agents and hard coding.
Faster V4 model: 284B total / 13B active, 1M context, Instant Mode on chat, economical API default.
Experimental V4-Flash that accepts image input (JPEG, PNG, GIF, WebP) alongside text on the API.
Reasoning-first predecessor: Thinking in tool-use, live on App, Web, and API when it shipped.
Hybrid Think / Non-Think generation, 128K context at launch, first step toward the agent era.
Reasoning model trained with large-scale RL; MIT weights; DeepThink on chat when it launched.
671B MoE / 37B active open model that powered the free app at launch.
I am not DeepSeek. Features, prices, and retirement dates change on their side. Confirm anything you will ship against the official DeepSeek website and the API pricing page. This overview is mine.
Product groups
Free chat at chat.deepseek.com with Instant Mode and Expert Mode.
OpenAI- and Anthropic-shaped API at api.deepseek.com for V4 models.
Local-first coding agent runtime in developer preview. Everything is a plugin.
Optional chain-of-thought before the answer; effort low, high, or max.
V4 and earlier checkpoints on Hugging Face under the project's open licenses.
Why I keep an API key
The homepage still offers Chat Now as free access. I wrote this article’s outline there before I moved stubborn passages to Expert Mode.
V4’s launch note is blunt: 1M context is now standard across official DeepSeek services. I stopped splitting RFCs into three chats.
OpenAI Chat Completions and Anthropic-shaped calls both work. I changed base_url and the model string. I did not rewrite a client.
The V4 note lists Claude Code, OpenClaw, OpenCode, and Codex. Harness exists for people who want the loop on their laptop.
Open the product, or stay with the model pages on this site.
Try nowWho this is for
If your stack already speaks OpenAI or Anthropic APIs, Flash and Pro are a model-name change. Start with DeepSeek-V4-Flash unless the job is a hard agent.
The app and web chat are free on the official doors. Instant Mode is enough for mail and notes. Expert Mode is the one I use when I care about the reasoning.
Harness is explicitly a developer preview for people building harnesses, not a polished consumer IDE. If that sentence makes you curious, you are the audience.
V4 preview was open-sourced with a technical report. Earlier V3 and R1 checkpoints remain on Hugging Face. License terms live on those cards, not on this page.
Questions I kept hearing
DeepSeek is Hangzhou DeepSeek's family of language models plus a free chat surface and a developer API. I use the official English homepage as the front door: chat for people, the API platform for apps, and Hugging Face for weights. This page is an independent overview. It is not the official DeepSeek site. Current chat and API traffic on the pages I opened is V4 — DeepSeek-V4-Pro and DeepSeek-V4-Flash — with an experimental vision variant on the API.
The homepage I resolved is the official DeepSeek website at www.deepseek.com/en/. Chat lives at chat.deepseek.com. The API platform is platform.deepseek.com, with docs at api-docs.deepseek.com. App downloads go through download.deepseek.com/app/. I do not treat social posts as a source of product truth.
On chat.deepseek.com the V4 announcement points Instant Mode at Flash and Expert Mode at the heavier Thinking path. Instant is the one I reach for when I want a first pass. Expert spends tokens on Thinking before it answers. Both sit on a 1M context window in the current official services. I do not treat the names as separate products — they are modes on the same V4 lineup.
The API lists deepseek-v4-flash-vision-exp as an experimental model that accepts image input — screenshots, photos, charts — in user messages. Formats on the docs I opened are JPEG, PNG, GIF, and WebP. That is reading pictures, not drawing them, and not video generation. I did not find a video-generation product on the official site, so this overview does not pretend one exists.
Harness is a local-first coding-agent runtime in developer preview. The official page's line is 'everything is a plugin': models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI. I launched the idea with npx @deepseek-ai/dsh web as their quick start. Source is on GitHub at deepseek-ai/deepseek-harness. It is not a replacement for the hosted chat tab.
No. This is an independent notes page. Product facts come from pages I opened on the official site and API docs. For accounts, keys, and downloads, use the official DeepSeek website. Primary buttons here open a local next-step dialog, not a login on DeepSeek.
Reader notes
Priya N.
★★★★★ 5 out of 5
I pointed Claude Code at DeepSeek-V4-Pro for a brownfield service. It found a race I had papered over, then wrote the test. I still merge the patch myself.
Marco T.
★★★★☆ 4 out of 5
Flash is the one I keep for outlines. Instant Mode is quick; I switch to Expert when the draft has to survive a design review.
Hannah L.
★★★★★ 5 out of 5
I pasted an RFC that used to need three chunked chats. The 1M window kept the appendix in play. Two 'open' questions were already closed — that was on me.
Wei C.
★★★★★ 5 out of 5
Swapping base_url to api.deepseek.com in the OpenAI SDK took minutes. Tool calls and JSON mode behaved as the docs described.
Elena V.
★★★★☆ 4 out of 5
Vision-exp read a messy ops chart and named the Tuesday dip. It does not draw pictures. I treat it as a reader, not a generator.
That split is the whole review. Open DeepSeek when you want to try it; stay on this site when you want the map.
Open Model catalog