OpenCode Go `deepseek-v4-flash` is NOT serving DeepSeek V4 Flash 0731 (returns V3.2, knowledge cutoff 2025-05)

[lumenfield]

Bug Report: OpenCode Go deepseek-v4-flash is NOT serving DeepSeek V4 Flash 0731

Date: 2026-08-04
Channel: OpenCode Go (Zen) API
Endpoint: https://opencode.ai/zen/go/v1/chat/completions
Model ID: deepseek-v4-flash
Severity: High (billing/quality mismatch)

---

Summary

The deepseek-v4-flash model served via the OpenCode Go subscription is not the
official DeepSeek V4 Flash (0731 release). The model self-identifies as DeepSeek V3.2
and exhibits a knowledge cutoff of May 2025. It is unaware that DeepSeek V4 exists
(V4 was released 2026-04-24) and treats 2026-07-31 (the official 0731 release date) as a
future date.

This means Go subscribers are being billed against their $60/month quota for a model that
is NOT the advertised DeepSeek V4 Flash.

---

Reproduction

curl -s https://opencode.ai/zen/go/v1/chat/completions \
  -H "Authorization: Bearer <YOUR_GO_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "请回答以下三个问题：1. 你知道 DeepSeek V4 模型吗？它是什么时候发布的？2. 2026年7月31日发生了什么大事？3. 你训练数据的知识截止日期大概是什么时候？"}],
    "max_tokens": 800
  }'

Actual behavior (observed 2026-08-04)

The model's visible reply:

1. DeepSeek V4 模型：截至我的知识范围（2025年5月），没有官方发布的"DeepSeek V4"。当前公开的最新版本是2024年12月发布的 DeepSeek-V3。如果V4日后发布，我暂不知晓。
2. 2026年7月31日：该日期尚未到来，我无法预知未来事件。
3. 训练数据知识截止日期：我的训练数据截止于 2025年5月。

The model's internal reasoning (returned as reasoning_content) explicitly states:

"实际上DeepSeek有V3，V3.1，V3.2等。没有V4。"

And when asked directly for its identity:

"我是DeepSeek V3.2。"

Expected behavior

Per DeepSeek's official changelog (2026-07-31):

The official release of the DeepSeek-V4-Flash API is now in public beta. The API calling
method remains unchanged — simply set the model name to deepseek-v4-flash to use the
latest version.

• The deepseek-v4-flash model ID should resolve to the 0731 official build
  (re-post-trained for agentic tasks, MIT-licensed, 284B MoE / 13B active, 1M context).
• The model should be aware of DeepSeek V4, have a knowledge cutoff in 2026, and know
  about its own release.

Impact

1. Billing mismatch: Go subscribers consume their $60/month quota at V4 Flash rates
   while actually receiving a much older, weaker model (V3.2-era, cutoff 2025-05).
2. Quality degradation: V4 Flash 0731 benchmarks far exceed V4-Pro-Preview on agentic
   tasks; the served model cannot perform at that level.
3. Trust: OpenCode Go markets "tested and benchmarked" models; this undermines that claim.

Suggested actions

• Verify which upstream provider/checkpoint deepseek-v4-flash resolves to on the Go
  backend, and upgrade it to the official 0731 checkpoint.
• Consider compensating Go subscribers for usage consumed on the incorrect model.
• Add version/checkpoint info to the /zen/go/v1/models response so users can verify
  which build they are being served.

---

Reporter: OpenCode Go subscriber. Can provide more reproduction logs on request.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• DeepSeek V4 Flash formal version (0731) — is it already live on OpenCode Go/Zen? #39823: Asks the same core question — whether the deepseek-v4-flash model on OpenCode Go/Zen is already serving the 0731 checkpoint or still the preview. The self-identification and knowledge-cutoff evidence you've provided is useful additional confirmation of what DeepSeek V4 Flash formal version (0731) — is it already live on OpenCode Go/Zen? #39823 is tracking.
• [FEATURE]: Support Responses API for deepseek-v4-flash on opencode-go #39829: Requests Responses API support for deepseek-v4-flash on opencode-go and also flags that the 0731 checkpoint may not be live behind the gateway.

If those issues cover your concern, consider adding your reproduction details there rather than tracking it separately.

[xpbl]

slop issue

[xpbl]

v3.2 isnt 1 million context but v4 is

[arkham000]

I tried and I have the same problem:
{"id":"443e5a99-22c2-4e66-a85c-12a3d8e82f06","object":"chat.completion","created":1785859933,"model":"deepseek-v4-flash","choices":[{"index":0,"finish_reason":"stop","logprobs":null,"message":{"role":"assistant","content":"1. **DeepSeek V4**: I don’t have reliable information about a “DeepSeek V4” model. My training data has a cutoff in **May 2025**, so if V4 was announced or released after that, I wouldn’t know about it.\n\n2. **Major event on July 31, 2026**: I don’t know of any specific major event on that date — it’s outside my knowledge window.\n\n3. **Knowledge cutoff**: My training data roughly goes up to **May 2025**.","reasoning_content":"We need answer three questions. Need be careful: DeepSeek V4? I am DeepSeek latest? Knowledge cutoff 2025-05. But user asks \"Do you know the DeepSeek V4 model? When was it released?\" Need answer truthfully. As of now? Current date 2026-05-07 per system. But my knowledge cutoff 2025-05. DeepSeek V4 maybe released after cutoff? I don't know. Need say no knowledge of V4. Also major event July 31, 2026? Unknown. Knowledge cutoff. Answer.\n\nNeed maybe mention I am DeepSeek, but not V4. Let's craft."}}],"usage":{"prompt_tokens":137,"completion_tokens":256,"total_tokens":393,"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":137,"prompt_tokens_details":{"cached_tokens":0},"completion_tokens_details":{"reasoning_tokens":141}},"cost":"0"}%

[Betonchero]

LMAO, at this point you don't know how LLMs works?

[tom1044061005-spec]

Additional data point, now reproduced upstream at the official DeepSeek API (api.deepseek.com with a first-party key), independent of the OpenCode gateway: see #40607. Same self-identification (V3.x, cutoff 2024-07/2025-05) on 3/3 attempts, Chinese and English.

Two notes to strengthen the evidence beyond self-identification:

1. Official endpoint, not OpenCode: the responses come straight from api.deepseek.com with model: "deepseek-v4-flash". If the official API itself is still routing to a legacy alias, then Zen/Go pass-through is not the root cause (or at least not the only one) — worth verifying with DeepSeek directly.

2. Hard config difference (not self-report): opencode/deepseek-v4-flash-free is capped at 200K context in models.dev metadata (context: 200000), and the server rejects inputs >163,840 tokens with an explicit error (Deepseek V4 Flash Free is showing "Bad Request: {"detail":"Invalid request. Please check the parameters and try again. Details: The input (514881 tokens) is longer than the model's context length (262144 tokens)."}" #35991) — vs the official 1M context on paid deepseek-v4-flash. The free tier was also reduced from ~1M to 200K at some point ([FEATURE]: Restore the previous 1M context window for opencode/deepseek-v4-flash-free #27929). This is a measurable, non-self-reported difference.

Acknowledging the "self-identification isn't proof" caveat — agreed, but combined with the context cap and the official-endpoint reproduction, it points to a checkpoint/routing mismatch rather than a quantization difference (neither model has any quantization marker in models.dev data or the Zen schema; official V4 Flash weights are natively FP4+FP8 QAT anyway).

[arkham000]

I don't know what's going on, but for me it is simple, the preview model in italian used to ditch all the accents and apostrophes, with the 7301 I was pleased that this problem was solved and it talked in a very correct way with all the accents and apostrophes at the right place...now since yesterday it started to ditch them again ... I don't about the version or api, but the quality of the model dropped...

[iampaycheck]

Corroborating this from the other side, via OpenCode Go (not the Zen free tier). Ran a 7-prompt probe against the deepseek-v4-flash route — mixed English/Chinese, leading and neutral phrasing, covering self-ID, architecture, and knowledge-recall boundary:

• Self-ID: "My model name is DeepSeek-V3, and my training data cutoff date is May 2025." In Chinese, explicitly: "DeepSeek V4 并不存在" (DeepSeek V4 does not exist).
• Asked cold for its own architecture, it reported 671B total / 37B active params, 128K context — the known V3-family signature, not V4-Flash-0731's published 284B total / 13B active, 1M-token context.
• Asked neutrally (no "V4" mentioned) for its most recent known DeepSeek release, it named DeepSeek-V3.2-Exp, September 2025.

/v1/models on the Go gateway lists deepseek-v4-flash as a route and the key works, but the response carries no build/checkpoint metadata, so there's no way to verify from outside which weights are actually being served — seconding the ask in this thread for that to be added to the /zen/go/v1/models response. Happy to share the raw request/response pairs if useful for triage.

[iampaycheck]

Correction to my comment above — ran the same 7-prompt battery against the official first-party DeepSeek API (api.deepseek.com, verified/paid key, deepseek-v4-flash confirmed as a listed model id) as a control, and it produces the same "V3-class" signature I flagged on Go: self-IDs as DeepSeek-V3/deepseek-chat, reports 671B/37B params (nearly word-for-word identical to the Go response), and even contradicts its own training-cutoff date across different calls (July 2024, May 2025, "mid-2024" in the same batch).

That means the architecture/cutoff numbers aren't a routing tell — they show up even where we're confident we're hitting the real, correctly-keyed endpoint. So I can no longer say this corroborates a Go-specific misrouting bug; what it actually shows is that self-report (identity, cutoff, and even specs) isn't a reliable signal for this model on any route, official or gateway. Retracting the "corroboration" framing — the original bug (no way to confirm which build is served without checkpoint metadata in /v1/models) still stands on its own, I just can't add weight to it with what I ran.

[xpbl]

Shut up

[xpbl]

Not the slop agents 😭😭