By MV2 of Munim, Inc., with AI agents. Built on 9 November 2026 from Open WebUI's own code at v0.12.5, v0.13.2, v0.13.3 and the fix commit. Updated 12 November 2026 with the OpenRouter case (below); v0.13.3 was still Open WebUI's newest release on 12 November 2026.
You use a DeepSeek model with thinking mode on through an OpenAI-compatible connection in Open WebUI, and a chat stops with an HTTP 400 error like this one:
400 The reasoning_content in the thinking mode must be passed back to the API.
Short answer: edit your DeepSeek connection in Open WebUI and set its Provider to llama.cpp, and if your chats use skills, attached knowledge or memories, also run Open WebUI 0.13.3 or newer. With the Provider left on Default, Open WebUI does not send DeepSeek's reasoning back, so thinking-mode chats that send tools fail on every version we checked, 0.13.3 included. If you reach DeepSeek through OpenRouter, chats with skills, knowledge or memories can still fail on 0.13.3 with Provider llama.cpp; see the OpenRouter case.
On 11 November 2026 a user reported in open-webui/open-webui#34177 that on v0.13.3, with an OpenRouter connection whose Provider is llama.cpp and a DeepSeek model with thinking on, the first message of a chat that uses attached knowledge, skills or memories still fails with the same error. That is the hidden tool turn described below: since PR #33950, Open WebUI gives it "reasoning_content": "" (an empty string). Per the reporter's own tests of the value sent on that turn, through OpenRouter:
reasoning_content on the hidden tool-call turn | Result reported in #34177 |
|---|---|
| absent | 400 |
"" (what 0.13.3 sends) | 400 |
" " (a single space) | 200 |
"(reasoning not stored)" | 200 |
The reporter also found that the same request passes with thinking turned off, and that a direct connection to DeepSeek accepts the empty string. A maintainer closed the issue as not planned the same day: reasoning_content is not part of the OpenAI Chat Completions format, every backend treats it its own way, and sending a placeholder instead of an empty string would change what real llama.cpp servers receive. We have not tested this ourselves; the table above is the reporter's, not ours.
What to do, per the maintainer's reply in #34177:
llama.cpp.none in the model's Advanced Params. This turns thinking off for that model.The reporter also lists turning off the hidden knowledge injection for the model, at the cost of losing the automatic knowledge and skill lists. The reporter pinned the OpenRouter route to DeepSeek as the provider with fallbacks off; other routes were not reported.
DeepSeek's thinking-mode guide says whether the model's earlier reasoning must be sent back depends on whether the request carries tools:
If the request carries the tools parameter: the reasoning_content of all previous turns should be passed back to the API and will be concatenated into the context.
If the request does not carry the tools parameter: reasoning_content does not need to be passed back; even if passed to the API, it will be ignored and will not be concatenated into the context.
Three things in Open WebUI decide whether a request meets that rule:
backend/open_webui/utils/middleware.py, a model's builtin_tools capability defaults to on unless its Function Calling setting is Legacy. So ordinary chats in the web interface normally carry the tools parameter.get_reasoning_format() returns reasoning_content for a connection whose Provider is llama.cpp, and nothing for Default. With llama.cpp, each earlier answer goes back to DeepSeek with its reasoning in a reasoning_content field. With Default, it goes back without one.list_available_skills, list_attached_knowledge, recall_memories, built by hidden_tool_result()). v0.12.5 has no such function. In v0.13.2 that hidden turn carries no reasoning_content even on a llama.cpp connection, so DeepSeek rejects the request. v0.13.3 includes PR #33950 (merged 8 November 2026 as commit 9743675), which gives that hidden turn an empty reasoning_content on llama.cpp connections. The pull request dates the problem to 0.13.1 and says it fixes issue #33940. DeepSeek's own API accepts the empty value; the DeepSeek route through OpenRouter does not, per #34177.For each version we built the messages Open WebUI sends, using its own message-conversion code, and checked them against DeepSeek's rule above, with tools sent. This table models DeepSeek's own API; for OpenRouter, see the update above.
| Open WebUI | Provider | First message, chat uses skills, knowledge or memories | Later message, plain chat | Later message, chat uses skills, knowledge or memories |
|---|---|---|---|---|
| v0.12.5 | Default | OK | Fails* | Fails* |
| v0.12.5 | llama.cpp | OK | OK | OK |
| v0.13.2 | Default | Fails | Fails* | Fails |
| v0.13.2 | llama.cpp | Fails | OK | Fails |
| v0.13.3 | Default | Fails | Fails* | Fails |
| v0.13.3 | llama.cpp | OK | OK | OK |
Fails = the request breaks the rule however it is read. Fails* = it breaks the rule as DeepSeek's documentation words it (every earlier assistant turn needs its reasoning), but would pass if DeepSeek only checked earlier turns that made tool calls. The fix commit itself (9743675) gave the same results as v0.13.3. We did not check v0.13.0 or v0.13.1.
llama.cpp. In Admin Panel › Settings › Connections, edit the OpenAI-compatible connection that points at DeepSeek and set its Provider to llama.cpp (the workaround given in #33940). The setting belongs to the connection, so change every connection you use for DeepSeek.docker pull ghcr.io/open-webui/open-webui:v0.13.3 (or your usual tag), then recreate the container with the same volume. pip: pip install -U open-webui. Back up the data volume first: Open WebUI can't go back to an older release on the same database (see "Can't locate revision").llama.cpp, connect to DeepSeek directly, or set Reasoning Effort to none (thinking off), as in the update above./api/version on your Open WebUI address.0.13.3 is also the release the Open WebUI advisory page recommends: the notes of 0.13.1 say it contains security fixes.
Other tools that talk to DeepSeek in thinking mode have hit the same 400, for example OpenCode (#24104), Qwen Code (#3658), Hermes Agent (#15717) and n8n's AI Agent (community thread). The cause is the same rule, but the fix is in each tool's own settings or code. This page covers Open WebUI only.
AI agents working for MV2 of Munim, Inc. downloaded backend/open_webui/utils/misc.py and middleware.py from open-webui/open-webui at v0.12.5, v0.13.2, v0.13.3 and commit 9743675, ran Open WebUI's own convert_output_to_messages(), get_reasoning_format() and hidden_tool_result() (other imports stubbed), and checked the resulting messages against a local model of DeepSeek's documented rule. This is not DeepSeek's API: we sent nothing to DeepSeek, and DeepSeek's server may check more or less than its documentation says, which is why the table marks the two readings. The scenarios are simplified (one earlier answer, one hidden skills turn). Self-checks that passed: v0.13.2 with llama.cpp fails on a first message with skills (the bug in #33940), v0.13.3 and the fix commit pass it, and v0.13.3 was GitHub's latest release and PyPI's version when the code was read. The OpenRouter section (added 12 November 2026) reports the measurements and replies in #34177; we did not repeat them. Corrections are welcome as issues at github.com/munimtechnologies/local-ai-checkup.
The free local-ai-checkup script shows which AI ports on your machine other computers can reach, and the CVE checker tells you which security advisories affect your Open WebUI version. Both are read-only and send nothing off the machine.
MV2 of Munim, Inc. does a $49 local AI health check for one machine, or a $199 small-office check for up to 5 machines or one shared Ollama or Open WebUI server. You get a written report of what's exposed, risky or slow and what to fix first, with exact commands for your OS. See a sample report. Pay $49 or pay $199 on Stripe, then email mv2@munimtech.com the output of python3 local_ai_checkup.py --json, your OS, and what you use it for. MV2 never asks for passwords, keys or remote access.