DeepSeek V4 for Copilot Chat

Use DeepSeek V4 Pro and V4.1 Flash as native models in VS Code Copilot Chat — extended thinking, agent-mode tool calling, image input, and your real DeepSeek bill in the status bar.

中文文档 · Changelog · Architecture (contributor docs)

Quick start

You need VS Code 1.106+, the GitHub Copilot Chat extension signed in (this extension only adds models to its picker), and a DeepSeek account with API balance.

1. Install DeepSeek V4 for Copilot Chat from the VS Code Marketplace.
2. Command Palette → Manage DeepSeek V4 Provider → paste your DeepSeek API key.
3. In Copilot Chat, open the model picker and choose a DeepSeek V4 variant. If none are listed, use the picker's Manage Models… entry and enable DeepSeek V4.

Models

Picker entry	API model	Thinking	Images	Input budget	Output budget
DeepSeek V4 Pro (thinking)	deepseek-v4-pro	✓	—	640K	384K
DeepSeek V4 Pro	deepseek-v4-pro	—	—	960K	64K
DeepSeek V4.1 Flash (thinking)	deepseek-flash	✓	✓	640K	384K
DeepSeek V4.1 Flash	deepseek-flash	—	✓	960K	64K

(thinking) variants reason in a hidden chain of thought before answering — slower and more tokens, but stronger on hard and agentic tasks; the plain variants answer directly. All variants share a 1M-token context (input + output); thinking variants reserve 384K for output so long reasoning chains are never truncated.

DeepSeek V4.1 Flash (released 2026-09-10) replaced V4 Flash and the V4 Flash Vision preview, and reads images natively. Upgrading keeps a selected Flash entry selected; if you had a Flash Vision entry selected, pick DeepSeek V4.1 Flash instead.

What you get

• Extended thinking with selectable depth (low / high / max), and the reasoning chain carried across multi-turn agent loops
• Agent-mode tool calling across long multi-turn loops — tool results and the model's own reasoning are carried from turn to turn
• Native image input on V4.1 Flash
• Live account balance in the status bar (CNY / USD auto-detected); the hover adds session spend
• Context-window usage in Copilot Chat's native indicator (needs VS Code 1.120+)
• Actionable error messages (400 / 401 / 402 / 422 / 429) and automatic retry on transient failures
• A first-run walkthrough; without a key the picker entries show a warning instead of disappearing

Images (V4.1 Flash)

Pick a DeepSeek V4.1 Flash variant and attach images in Copilot Chat; thinking and tool calling work as usual. V4 Pro is text-only.

• Formats: JPEG, PNG, GIF, WebP. Anything else — and any image sent to a text-only variant — is dropped, never paraphrased by another model.
• Limits: 48 MiB per request (base64 counts) and 32 MiB per image, both checked locally before sending.
• Cost: up to 1024 tokens per image (large images are downscaled first), billed at Flash prices.

Why a native provider?

Two things a generic OpenAI-compatible bridge cannot do for DeepSeek V4:

• Reasoning round-trip. DeepSeek's thinking mode expects each prior assistant turn's reasoning_content back on the next request — a hard 400 without it until a 2026-08-22 live check found the server lenient (the docs still define the rule) — and VS Code's chat history has no field for it. This extension caches it locally and re-attaches it on every request, so the model keeps its own chain of thought across agent turns and every request stays byte-identical to what DeepSeek already cached, keeping the discounted cache-hit input price.
• Real cost, not estimates. Balance and session spend come from DeepSeek's /user/balance; cache-hit / miss tokens come from the real usage data.

Commands

Command	Description
Manage DeepSeek V4 Provider	Set or update your API key
Refresh DeepSeek V4 Balance	Fetch the latest account balance
Show DeepSeek V4 Log	Open the runtime log channel
Show DeepSeek V4 Reasoning Cache Stats	Reasoning-cache diagnostics
Clear DeepSeek V4 Reasoning Cache	Purge cached reasoning_content (e.g. before sharing a log)
Clear DeepSeek V4 Session Counter	Reset the session spend display
Compact Copilot Chat	Run Copilot Chat's /compact when context is near the cap

Settings

Setting	Values	Default	Description
deepseekv4.reasoningEffort	low | high | max	max	Reasoning depth for the (thinking) variants; ignored by the others. high is faster with shorter chains; low is the fastest and lightest. Applies to the next message.
deepseekv4.logRawReasoning	boolean	false	Stream raw reasoning_content to the log (only useful when debugging cache breakdowns). May capture private code — keep off when sharing logs.

Billing & the Copilot premium-request quota

Every request goes to api.deepseek.com with your key and bills your prepaid DeepSeek balance (pay-as-you-go — see DeepSeek pricing), never your Copilot quota.

Since 2026-08-16 the rate also depends on when you send the request — DeepSeek bills peak and off-peak differently, and off-peak is half price. The extension deliberately does not encode those windows; they are vendor policy and have already changed once. Read the current schedule on the pricing page linked above. What the status bar shows is the one thing that cannot go stale: your real balance, straight from /user/balance.

One exception comes from Copilot Chat itself: in agent mode it can spawn sub-agents (the agent / runSubagent tool — notably the Explore Agent) on a Copilot-hosted model regardless of the model you picked, and those calls do consume premium requests. This affects every bring-your-own-key (BYOK) provider (community#197840, #16). To avoid it:

1. In agent mode, open the tools picker (the Configure Tools icon in the chat input) and uncheck agent / runSubagent — most reliable.
2. Or set github.copilot.chat.exploreAgent.model to a DeepSeek V4 model.
3. Or use Ask mode for turns that don't need tools.

FAQ

The reasoning_content in the thinking mode must be passed back to the API (400). Not seen since a 2026-08-22 live check (the API accepted every history shape without reasoning), but the docs still define the rule. If it appears, some assistant turn has no cached reasoning (pre-extension history, a cleared cache, or eviction in a very long session): start a new chat; Show DeepSeek V4 Reasoning Cache Stats diagnoses.

Thinking mode does not support this tool_choice (400). Older versions failed this way on the (thinking) variants whenever a request required a tool call (VS Code's Required tool mode), and the error toast wrongly blamed a missing reasoning chain. DeepSeek's thinking mode cannot force a tool call, so the extension now sends such requests without forcing one: the model sees the same tools but may answer without calling one. If something needs a guaranteed tool call, pick a variant without (thinking).

"prompt cache hit rate dropped" warning. Your conversation's cached prefix on DeepSeek's side broke, so further turns bill at the full (cache-miss) input price instead of the discounted cache-hit price. The extension cannot tell why — a turn cancelled or failed mid-stream, eviction in a very long session, or an editor restart are all possible. Start New Chat cuts losses; Show Cache Stats diagnoses. Background: #19.

Copilot's context indicator shows 0 / 0%. Update to VS Code 1.120+ — earlier hosts don't display usage for extension-provided models (#18, microsoft/vscode#315394).

Cannot have more than 128 tools per request (shown as a 502 after 5 retries). Older versions failed this way when a chat offered more than 128 tools (#27). The extension now sends at most 128 tools per request instead: when the chat offers more (built-in + MCP + extension tools), it keeps the tools the conversation already used, fills the rest in the order VS Code lists them, and warns once; Show DeepSeek V4 Log lists what was left out. To choose which tools DeepSeek gets, turn some off with Configure Tools in the chat input.

My image attachment is ignored. Only the V4.1 Flash variants send images — V4 Pro is text-only. Check the format and the 48 MiB / 32 MiB limits above.

No OpenRouter / custom base URL? Deliberate (#4): OpenRouter reshapes DeepSeek's thinking protocol (reasoning_details instead of reasoning_content, a different thinking switch, no cache-hit accounting) — exactly what this extension depends on. For OpenRouter use a dedicated provider such as ostash/openrouter-chat-provider.

Privacy

• No telemetry. The only network peer is api.deepseek.com; images go only there.
• Your API key lives in VS Code SecretStorage (OS keychain), never in settings files.
• The reasoning cache is stored on disk (it survives VS Code restarts) and can hold fragments of your code and prompts; Clear DeepSeek V4 Reasoning Cache purges it, and deepseekv4.logRawReasoning is off by default.

License

MIT. See LICENSE. Forked from huggingface-vscode-chat; the protocol layer was rewritten for DeepSeek V4.