Skip to content

v2.0.0: DeepSeek V4

Choose a tag to compare

View all tags
@arikusi arikusi released this 22 Jun 16:34
v2.0.0
c06a79b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

DeepSeek V4

This release migrates the server and the hosted endpoint to DeepSeek V4.

Models

  • deepseek-v4-flash (new default): fast and economical, 1M context, up to 384K output
  • deepseek-v4-pro: top capability, 1M context, up to 384K output
  • deepseek-chat and deepseek-reasoner keep working as aliases that resolve to deepseek-v4-flash (chat = non-thinking, reasoner = thinking). The DeepSeek API retires those two names on 2026-07-24, so the server translates them for you. Upgrading is drop-in.

Highlights

  • Default model is now deepseek-v4-flash. Direct v4 calls are non-thinking by default for fast responses; enable reasoning with thinking: {type: "enabled"} or the deepseek-reasoner alias.
  • New reasoning_effort parameter (high / max) for thinking mode.
  • max_tokens upper bound raised to 384000.
  • Model fallback now pairs deepseek-v4-flash with deepseek-v4-pro.
  • The hosted worker endpoint (deepseek-mcp.tahirl.com) is on V4 as well.

Pricing (per 1M tokens)

  • v4-flash: $0.0028 cache hit, $0.14 cache miss, $0.28 output
  • v4-pro: $0.003625 cache hit, $0.435 cache miss, $0.87 output

Fixed

  • Declared cost_usd and routed_from in the deepseek_chat output schema, which strict MCP clients rejected under SDK 1.29 structured-content validation.

Full notes in CHANGELOG.md.

Assets 2
Loading