Misc. bug: Qwen3.5 with Opencode: Assistant response prefill is incompatible with enable_thinking.

Name and Version

version: 8468 (3306dba)
built with Clang 19.1.5 for Windows x86_64

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

No response

Command line

Problem description & steps to reproduce

when using OpenCode with a Qwen3.5 model with thinking enabled, this error sometimes appear causing the request to fail.

I've tried various combinations of models paramters, llama.cpp settings nad Opencode config with no luck.

any recommendation? My understanding from issues is that this should have been fixed a while ago, but I have been experiencing this error since the release of Qwen3.5 and it happens even on latest build.

First Bad Commit

No response

Relevant log output

Logs

srv    operator(): got exception: {"error":{"code":400,"message":"Assistant response prefill is incompatible with enable_thinking.","type":"invalid_request_error"}}

This issue might be similar or related to the following issue(s):

Misc. bug: enable_thinking param cannot turn off thinking for qwen3.5 #20182: enable_thinking param cannot turn off thinking for qwen3.5
Eval bug: Bug: Qwen3.5 enable_thinking=false via --chat-template-kwargs is ignored across all shells (PowerShell/Bash) #20409: Bug: Qwen3.5 enable_thinking=false via --chat-template-kwargs is ignored
Eval bug: Qwen3.5, when resuming an interrupted completion, <think> tags are inserted into the output token stream #20768: Qwen3.5, when resuming an interrupted completion, tags are inserted into output
Eval bug: Qwen3.5 9B often prints tool calls in XML and stops when thinking is enabled - tool calls inside thinking block #20837: Qwen3.5 9B often prints tool calls in XML and stops when thinking is enabled
Eval bug: Response always starts with </think> tag when running Qwen3.5 9B #20516: Response always starts with tag when running Qwen3.5 9B
Eval bug: llama-server (b8400): ngram parameters corrupted to 1024 and <think> tags forcefully injected despite reasoning_format=none #20789: ngram parameters corrupted and tags forcefully injected despite reasoning_format=none
Eval bug: llama-cli crash in common_chat_peg_parse (std::runtime_error) with Qwen 3.5 Thinking model #19869: llama-cli crash in common_chat_peg_parse with Qwen 3.5 Thinking model
Eval bug: llama.cpp b8429: Qwen3-Instruct-2507 tool calling broken by false thinking detection #20809: Qwen3-Instruct-2507 tool calling broken by false thinking detection (CLOSED)
Eval bug: Qwen3-30B-A3B-Instruct-2507 displays incorrect thinking sections #20550: Qwen3-30B-A3B-Instruct-2507 displays incorrect thinking sections (CLOSED)
Eval bug: Thinking model with thinking disabled cannot use /apply-template with final assistant turn #15401: Thinking model with thinking disabled cannot use /apply-template with final assistant turn (CLOSED)

This comment was auto-generated locally using llama.cpp version: 8456 (b31b30f) on NVIDIA DGX Spark

Just add --no-prefill-assistant to avoid this issue. It's because the opencode didn't receive full response from AI provider. It sent incomplete assistant message back again and expect AI could continue to generate the rest of response. That make thinking mode model couldn't handle properly. Just use the --no-prefill-assistant as a workaround until there's a fix from llama or opencode.

If opencode wasn't receiving a full response from the provider (i.e. llama.cpp server), it seems like that would indicate a bug in llama.cpp.

It also only seems to occur for me with Qwen3.5 (27b) with a different agentic framework. I don't seem to get the issue with any other model but it is very intermitent. Next time I encounter this, I'll try and provide logs but just wanted to one up the issue.

Using a custom Jinja template worked for me. See:
#20837 (comment)

I also hit that today with version:
version: 8763 (ff5ef82)
built with GNU 11.4.0 for Linux x86_64
and Qwen 3.6 35B A3B

IDE: Opencode 1.14.18
Model: Qwen3.6 35B A3B IQ4_XS (Bartowski)
Server: llama.cpp b8850

Strangely enough, I encountered the bug twice today even with thinking disabled.

"Assistant response prefill is incompatible with enable_thinking"

Same,

 {"error":{"code":400,"message":"Assistant response prefill is incompatible with enable_thinking.","type":"invalid_request_error"}}

Parameters:

llama-server --hf-repo unsloth/Qwen3.6-35B-A3B-GGUF:Q6_K_XL --host 0.0.0.0 --port 11434 --ctx-size 256000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00

Confirmed : [https://github.com//issues/20861#issuecomment-4106288368] work for us

@janvitos

+1, same setup as @janvitos (effectively). Will update with details shortly.

Happens in Qwen3.6-35B as well.

Reproduced here too, Qwen3.6-35B-A3B (unsloth).

Request:

{
    "model": "qwen3.6-35b-a3b",
    "messages": [
        {
            "role": "user",
            "content": "Question: What are the names of the planets in the Solar System?"
        },
        {
            "role": "assistant",
            "content": "Answer: "
        }
    ]
}

Response:

{
    "error": {
        "code": 400,
        "message": "Assistant response prefill is incompatible with enable_thinking.",
        "type": "invalid_request_error"
    }
}

Interestingly, I tried it with Qwen3.5-4B with {"enable_thinking": false} (to make sure reasoning was off), and the content of the response started with <think>\n\n</think>\n\n as if the template wasn't removing it and wasn't incorporating the assistant prefill.

version: 9110 (ef22b3e)

This issue was closed because it has been inactive for 14 days since being marked as stale.

Navigation Menu

Repository navigation

Misc. bug: Qwen3.5 with Opencode: Assistant response prefill is incompatible with enable_thinking. #20861

Name and Version

Operating systems

Which llama.cpp modules do you know to be affected?

Command line

Problem description & steps to reproduce

First Bad Commit

Relevant log output

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Misc. bug: Qwen3.5 with Opencode: Assistant response prefill is incompatible with enable_thinking. #20861

Description

Name and Version

Operating systems

Which llama.cpp modules do you know to be affected?

Command line

Problem description & steps to reproduce

First Bad Commit

Relevant log output

Activity

github-actions commented on Mar 22, 2026

UltraRabbit commented on Mar 22, 2026

joshvoigts commented on Apr 4, 2026

marciowb commented on Apr 16, 2026

adrianhoehne commented on Apr 18, 2026

janvitos commented on Apr 19, 2026

gxcreator commented on Apr 20, 2026

sebastienroul commented on Apr 22, 2026

blockfeed commented on Apr 28, 2026

jhemmond commented on Apr 30, 2026

rneilson-cerby commented on May 13, 2026

github-actions commented on Jun 28, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions