Are there any open models that can actually compete with proprietary ones like GPT 5.5 Extended Thinking or Claude Opus 4.7? I am getting really good results with those in their chat interfaces for coding tasks. They sometimes spend 30-45 minutes working on my task and have an internal container they are doing tool calls on, like cloning a repository and compiling their code, and can find online documentation. Their answers are very good and usually correct for very complex tasks requiring specific protocols.

So I would like to know how well we can replicate this using open models since I want more control over how it runs, and privacy. Do any of you hook in agentic capabilities into your local models? How do you do it, and which models give you good results?

Pretend I have unlimited resources (local llama.cpp, sufficient fast storage/memory, and unlimited time to wait for a good response).

  • Zikeji@programming.devEnglish
    7·
    28 days ago

    I’ve been running Qwen 3.5 122B A10B but recently swapped to Qwen 3.6 35B A3B - both using OpenCode e as my agentic harness (though I’ve also used Pi). I’ve been happy with the output, though I have to be more precise with my prompts and do planning passes.

    • hok@lemmy.dbzer0.comOPEnglish
      2·
      27 days ago

      Thank you for your opinion & recommendations. Something I saw today related to “sub-agents” is in Kimi 2.6’s model card it says

      Elevated Agent Swarm: Scaling horizontally to 300 sub-agents executing 4,000 coordinated steps, K2.6 can dynamically decompose tasks into parallel, domain-specialized subtasks, delivering end-to-end outputs from documents to websites to spreadsheets in a single autonomous run.

      So maybe Kimi 2.6 is doing the “type of thing” I am looking for, but I don’t have the means to run it practically. Maybe at 1 token per second which would be brutal.

      I tried out Qwen 3.6 27B but not yet in an agentic setting, so I can’t really judge yet. Maybe it’s just me but the small model size seems limiting. I thought gpt-oss-120b was good.

  • cecilkorik@piefed.caEnglish
    2·
    27 days ago

    “Compete with”? Depends on your definition of competition. You can accomplish similar results with smaller, local models but you cannot do it as lackadaisically as with Claude by just throwing a sentence at it and letting it go to town for 15 minutes.

    Doing things like this locally will take more time and effort in countless different ways. You need to structure the prompts and the environment much more carefully. You need to wait much longer for much smaller portions of work. You need to retry again when it gets it wrong, which will happen, either relying on better luck or adjusting your plan, your prompts or your context to better guide it to what you’re actually looking for.

    If you’re used to Claude, working the same way with both and comparing them directly side by side, then no. Open models are not directly competitive like that. They can compete with it, if you’re willing to be much more involved in the process.

    If Claude is like a junior developer with access to an entire library of programming books, open models are like a 14-year-old in their first programming class with access to an entire library of programming books that they don’t know how to utilize effectively. They require a lot more guidance.

    You may wonder “what’s the point if I have to do so much work anyway, maybe I should just do it myself” and indeed, this is the crux of the problem. It’s even more obvious with smaller, open models than it is with the commercial AI models. This is not a new problem, it has been a problem even when training new employees. The difference is, real junior developers actually learn and grow based on my efforts to guide them and they eventually become senior developers. I’m not convinced that Claude or any open model ever actually will, despite how much effort goes into “training” them.

    • ɔiƚoxɘup@sh.itjust.worksEnglish
      3·
      28 days ago

      That’s my problem. None are there yet, at least with my hardware.

      If you’ve got 20 grand to spend, there’s a couple of models out there, like the one mentioned above that should do fine.

      • hok@lemmy.dbzer0.comOPEnglish
        3·
        28 days ago

        What I have yet to learn is how much of the intelligence and accuracy comes from the model itself and how much comes from the agentic tool system. For example, my experience with ChatGPT probably would be much worse with the free version (no thinking or container).

        • ɔiƚoxɘup@sh.itjust.worksEnglish
          1·
          27 days ago

          I’d say it’s 60-40 or 40-60. Both are important and have a large bearing on your results, but 128B and 8B will always have a big difference in reasoning capacity

  • troed@fedia.io
    2·
    27 days ago

    I run a quant of Qwen 35B A3B (Qwen3.6-35B-A3B-GGUF:UD_Q4_K_XL) at the moment, using Opencode and llama.cpp. I’m getting useful work out of it - but it’s of course not Claude. My hardware is a 5060Ti with 16GB VRAM and then ~20GB or so of system mem is getting used as well.

    It’s important to put boundaries on less capable models though, so I have two plugins in Opencode as well that really makes a big difference to the results: @tarquinen/opencode-dcp@latest and superpowers@git+https://github.com/obra/superpowers.git.

    I want to work in small steps with good control over what the models do so it’s not very similar to what you describe with just having them run away for half an hour and do everything.