Official DeepSeek API deepseek-v4-flash self-identifies as V3.x (cutoff 2024-07/2025-05), unaware of V4 Flash — reproduced with first-party key #40607

Description

Bug

Calling the official DeepSeek API (api.deepseek.com, first-party key) with model: "deepseek-v4-flash" returns a model that self-identifies as V3/V3.2/deepseek-chat with a knowledge cutoff of July 2024 / May 2025, and has no knowledge of DeepSeek V4 / V4 Flash (released 2026-04-24). The response model field says deepseek-v4-flash, but the actual served checkpoint appears to be a legacy V3.x alias.

This matches #40409 but is now reproduced upstream at DeepSeek's official API, not through the OpenCode gateway — suggesting OpenCode Zen/Go may be passing through a backend that is not yet serving the official V4 Flash (0731) checkpoint.

Observed responses (2026-08-05, 3 attempts)

• Chinese prompt: "我是 DeepSeek V3.2（API 版）…知识截止 2025 年 5 月…目前我没有 V4 系列的官方详细参数"
• English prompt: "I'm the DeepSeek assistant served via the API as deepseek-chat. My knowledge cutoff is July 2024. I have no knowledge of 'DeepSeek V4 Flash'"
• English repeat: "I am DeepSeek-V3, with a knowledge cutoff of July 2024. I have no information about DeepSeek V4 Flash or V3.2"

Related: free tier (opencode/deepseek-v4-flash-free) behaves the same

Independent opencode run instance: self-identifies as opencode/deepseek-v4-flash-free, no knowledge of V4 Flash / V3.2. Its context is capped at 200K (models.dev cache: context: 200000; server rejects >163,840 per #35991) vs 1M on the paid deepseek-v4-flash — reduced from ~1M per #27929.

No quantization marker exists for either model in models.dev data or the Zen model schema (no quant field). DeepSeek's official V4 Flash weights ship natively FP4+FP8 (QAT), so "quantized vs not" is not the actual difference.

Plugins

oh-my-openagent (only for the opencode run probe; official API test used plain curl/Invoke-RestMethod)

OpenCode version

1.17.11

Steps to reproduce

1. Create DeepSeek platform API key at platform.deepseek.com
2. curl https://api.deepseek.com/chat/completions -H "Authorization: Bearer $KEY" -d '{"model": "deepseek-v4-flash", "messages": [{"role": "user", "content": "What model are you? What is your knowledge cutoff? Do you know DeepSeek V4 Flash?"}]}'
3. Observe: response model field = deepseek-v4-flash, but content self-identifies as V3.x / deepseek-chat, cutoff 2024-07, no V4 knowledge
4. Repeat in Chinese and English — same result

Screenshot and/or share link

#40409

Operating System

Windows 11

Terminal

PowerShell 5.1