FIM in VIM with streaming?

[magikRUKKOLA]

How to do this in vim while using FIM?

high

@ubergarm @Thireus
What do you guys use?

[ubergarm]

@magikRUKKOLA

Heya, looks like you've vibe coded your own vibe coding client if I understand this correctly? I dont' know what FIM is and web search didn't give me a clear answer? Any context would be helpful for my overloaded splattered brain haha...

As for what do I use? I've used a mix of vim and neovim forever now... currently on neovim with LSPs but I don't have any LLM plugins or anything...

For vibe coding, I don't really do it. I've dabbled a bit and here are some very quick thoughts:

1. opencode - works okay, just need to setup an opencode.json. Thought some tool calls are wonky sometimes with Qwen35 but I think that was recently fixed so no more looping hopefully.
2. https://pi.dev/ some people say this is good because it is more light weight
3. i saw another one on reddit a couple days ago claiming to improve qwen3.5-35b-a3b performance with only a handful of custom tool calls especially reminding the LLM to check its work after every write_file call.. can't find it now though of course lol

Is high your own client then? Is your goal to integrate it with vim?

[magikRUKKOLA]

@ubergarm

Heya, looks like you've vibe coded your own vibe coding client if I understand this correctly?

Its more like a chat client with syntax highlighting and streaming.
I have used the mods for some time but then faced that it slows the actual LLM output so yeah, I made a clone of mods but without the bugs.

I dont' know what FIM is and web search didn't give me a clear answer?

Fill-In-the-Middle capabilities of the LLM. Some of them, like Qwen3.5-MoE-397B do support that. That is, in the vim or whatever one should be able to select that "middle" and hit a hotkey. Right after that the LLM should fill it. So I am looking for the exact same thing as I implemented in high -- I want the LLM output in vim to be outputted and highlighted with the streaming support.

[magikRUKKOLA]

@ubergarm

1. opencode

Uh oh. I seen the code, its a very bad software with a bunch of dependencies etc.
I already have the MCP server written in nginx-lua for a tool calling so I don't really need the functionality of opencode etc.
I also don't understand why there is such thing as Claude Code, opencode, openclaw etc. I am using vim. I don't feel like I need anything more than the FIM with streaming support for the vibecoding etc. So basically I am looking for a decent plugin with the same functionality as high but for vim (or even emacs for that matter). I wasn't able to find it.

https://pi.dev/ some people say this is good because it is more light weight

Does it support FIM? Does it support syntax highlighting with streaming? I don't think so.

i saw another one on reddit a couple days ago claiming to improve qwen3.5-35b-a3b performance with only a handful of custom tool calls especially reminding the LLM to check its work after every write_file call.. can't find it now though of course lol

Very interesting, but I never had such problems with the tool calls etc. Qwen3.5-397B works well with bash alone.

[magikRUKKOLA]

I dont' know what FIM is and web search didn't give me a clear answer?

https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/tokenizer_config.json

        "248060": {
            "content": "<|fim_prefix|>",
            "lstrip": false,
            "normalized": false,
            "rstrip": false,
            "single_word": false,
            "special": false
        },
        "248061": {
            "content": "<|fim_middle|>",
            "lstrip": false,
            "normalized": false,
            "rstrip": false,
            "single_word": false,
            "special": false
        },
        "248062": {
            "content": "<|fim_suffix|>",
            "lstrip": false,
            "normalized": false,
            "rstrip": false,
            "single_word": false,
            "special": false
        },
        "248063": {
            "content": "<|fim_pad|>",
            "lstrip": false,
            "normalized": false,
            "rstrip": false,
            "single_word": false,
            "special": false
        },

[ubergarm]

@magikRUKKOLA

Thanks for explaining, I believe I understand better what you want. Fill in the middle makes sense and I see how having that right inside of your preferred text editor (e.g. vim) would be good.

I also don't understand why there is such thing as Claude Code, opencode, openclaw etc. I am using vim. I don't feel like I need anything more than the FIM with streaming support for the vibecoding etc.

I felt this way too until a couple weeks ago, but now what I do is run opencode (or use whatever you want they are all very similar) in one terminal. Then I have my usual vim open in another terminal.

Then I manualy edit AGENTS.md or make up a file like PROMPT.md and edit that with vim. I then tell opencode to read that file and follow the instructions. I only edit files with vim natively while it does stuff under the hood as it likes. This allows it to make webfetch calls, use multiple tools, etc.

I combine that with git and if I like the results i git commit, if not I just undo the changes that it made. wash, rinse repeat...

---

Very interesting, but I never had such problems with the tool calls etc. Qwen3.5-397B works well with bash alone.

It isn't because of tool calling problems, they claimed it made Qwen3.5-35B-A3B score much better on some benchmarks by just reminding it to check its work.. I

okay i finally found the post: https://www.reddit.com/r/LocalLLaMA/comments/1rkdlqi/qwen3535ba3b_hits_378_on_swebench_verified_hard/

like everything on reddit, take it with a bucket of salt haha...

[magikRUKKOLA]

@ubergarm

I only edit files with vim natively while it does stuff under the hood as it likes. This allows it to make webfetch calls, use multiple tools, etc.

Yeah, I got your flow but I [would rather] prefer [to] closely monitor what the LLM is doing so I can interrupt and add the clarifications. The LLMs do a lot of stuff incorrectly etc. So there is no way I can delegate the decision-making to the LLMs themselves. :)

It isn't because of tool calling problems, they claimed it made Qwen3.5-35B-A3B score much better on some benchmarks by just reminding it to check its work.. I

Yeah, I do similar things all the time (as described above with the interruptions etc.), that's the exact reason I need highlighting and streaming -- to understand the code better and not to waste any time.

Thanks for the info!