Repository navigation
deepseek V4 flash connot close thinking mode #776
Description
Summary
deepseek V4 flash connot close thinking mode
Expected Behavior
{"thinking":{"type":"disabled"}}
or
{"reasoning_effort":"none"}
can enable
Actual Behavior
enable_thinking:false、reasoning:false will be ignore
and reasoning_effort:"none"：CommandCode response 400，only permit low/medium/high/xhigh/max。
Steps to reproduce the issue
enable_thinking:false、reasoning:false will be ignore
and reasoning_effort:"none"：CommandCode response 400，only permit low/medium/high/xhigh/max。
Command Code Version
1.38.2
Operating System
macOS
Terminal/IDE
Terminal
Shell
zsh
Session file (optional)
No response
Fix prompt (optional)
No response
Additional context
OS: macOS Sequoia 15.6
Metadata
Metadata
Assignees
Labels
No labels
Activity
EighTwT2 commentedon Sep 1, 2026
I can confirm the same issue. This does not appear to be client-specific.
Environment:
POST /provider/v1/chat/completionsdeepseek/deepseek-v4-flashResults:
{ "thinking": { "type": "disabled" } } The request returned HTTP 200, but thinking was still enabled. The response used all 2400 completion tokens as reasoning tokens and returned no visible content. 2. With: { "enable_thinking": false } The request also returned HTTP 200, but thinking was still enabled and reasoning tokens were still generated. 3. With: { "reasoning_effort": "none" } The API returned HTTP 400: Invalid option: expected one of "low" | "medium" | "high" | "xhigh" | "max" 4. Sending both reasoning_effort: "none" and thinking.type: "disabled" also returned HTTP 400. 5. reasoning_effort: "low" is accepted, but it does not disable thinking. In one test, the response still contained 1736 reasoning tokens. I also inspected the official command-code CLI package (1.39.2). It exposes reasoning effort levels but does not provide a supported off option. Could you please confirm whether the CommandCode DeepSeek V4 Flash route currently forces reasoning mode? If disabling reasoning is supported, what is the exact request field or model ID that should be used? No API keys, prompts, generated content, or private request data are included in this report.Hangsiin commentedon Sep 2, 2026
plz fix this.
sakura-madoromi commentedon Sep 5, 2026
same
ahmadbilaldev commentedon Sep 7, 2026
Working on the disable support.
a404670637-ch commentedon Sep 9, 2026
Post-fix retest (2026-09-09, Provider API): thinking still cannot be disabled or lowered below
low, andlowstill emits substantial reasoning tokens on several modelsFollow-up data after the fix announced in #697 (2026-09-07, "I have shipped a fix for this one. Please retry") — independent benchmark across 10 models via
POST https://api.commandcode.ai/provider/v1/chat/completions(OpenAI format, streaming, prompt:Reply with exactly: OK,max_tokens: 256, 1 warmup + 5 measured runs per model, sequential, 50/50 requests succeeded). Windows 10, direct HTTP, no CLI involved.1. There is still no off / minimal level. On
Qwen/Qwen3.7-Flash:{ "reasoning_effort": "minimal" }→
400Invalid option: expected one of "low"|"medium"|"high"|"xhigh"|"max"— unchanged post-fix.lowis the floor.2. Reasoning tokens actually emitted at
reasoning_effort: "low"(median of 5 runs, from the final usage chunk):Qwen/Qwen3.7-Flashmeta/muse-spark-1.2-contributorstepfun/Step-3.5-Flashtencent/hy3-paidQwen/Qwen3.8-Flashxiaomi/mimo-v2.5meituan/LongCat-2.0:freedeepseek/deepseek-v4-flash-fastz-ai/glm-5.3-flashpoolside/laguna-s-2.1-freeSo even at the lowest supported effort, a trivial one-line prompt still burns 120–180 reasoning tokens on Qwen 3.7 Flash / Muse Spark 1.2 Contributor / Step 3.5 Flash. Only
z-ai/glm-5.3-flashandpoolside/laguna-s-2.1-freeactually emit zero reasoning tokens atlow.Request: either expose a true disable (
reasoning_effort: "none"/"minimal", or forwardthinking: {"type": "disabled"}to upstreams that support it natively), or document the per-model minimum-effort behavior so agents can stop paying the latency/token tax. Happy to share the benchmark script and raw per-request JSON if useful.Clivos commentedon Sep 15, 2026
same
silvertakana commentedon Sep 18, 2026
Retested today and it still isn't fixed. This can't be that hard to fix..
TigerBeanst commentedon Sep 25, 2026
Still doesn't work on deepseek-v4.1-flash
Clivos commentedon Sep 27, 2026
It’s hard to imagine that an issue like this, opened a month ago, still hasn’t been resolved.
chanmankong commentedon Sep 28, 2026
same issue here. still not fix..
Hangsiin commentedon Sep 28, 2026
OpenCode Go handles this correctly, and this alone is a major reason for me to choose OpenCode Go. Please fix this issue.
ahmadbilaldev commentedon Oct 1, 2026
You can now turn off thinking on all deepseek models from /effort picker. Update to command-code@1.73.3+
Oyu-NAna commentedon Oct 2, 2026
Confirmed: setting "reasoning_effort": "off" successfully disables thinking