sorted by: new top controversial old
Qwen3.8-Flash-Next sounds like Claude and I hate it in c/localllama@sh.itjust.works
[–] 5 points 2 months ago

Never used Claude, but GPT-5.6 Sol annoyed me so much I actively avoid it and just use Terra on max. Sol is trained to favor convoluted San Francisco jargon that projects any trivial idea into some sort of an ingenious business maneuver.

  • permalink
  • fedilink
  • source
  • parent
  • context
    • 18
    Have you tried deepseek harness? (programming.dev)
     

    Yesterday I installed the deepseek harness after hearing a lot of noise about it from different channels. It's minimal but with the idea that everything is a plugin and agents can create plugins dynamically so it basically has infinite flexibility.

    Today I decided to get it to build a human/agent collaboration pipeline. For each workspace (project), an AGENTS.kanban is created that has a kanban board formatted as json to match the vscode extension portable-kanban. The board is fully viewable and functional in the harness web ui (as well as vscode) with an identical set of features to portable-kanban. Then I can use the command: /kanban next to instruct the agent to pickup the next todo task from the Doing or To Do boards and implement it.

    To implement all this I used Ornith-1.5-35B running on my AI PC (with a modest 32+8 memory), and when there were complex issues or the harness was throwing errors, I enlisted the help of GLM-5.3-Flash through the lite qwen token plan ($6!). I ran into two issues that caused previous dsh sessions to fail to parse both were debugged and fixed by GLM.

    Now I'm at end of day, I have a fully functional kanban collab plugin which I used to develop the plugin itself. The agent also started contributing new tasks into the backlog (ones marked with "Confirm" label).

    I'm not sure where I'll go next with it, but it's already usable! which is astonishing.

    My experience with deepseek harness so far:

    The Good

    • Coming from opencode, dsh is superior in its clarity and runtime loop
    • It's noticeably lighter to run than opencode on my system despite my running it through the web interface (dsh web) rather than TUI
    • The everything is a plugin philosophy is a great approach imo
    • Using workspaces as a first class citizen, workspaces map perfectly to projects and all are active and accessible always. So I can launch tasks on 2+ projects concurrently from the same interface

    The Bad

    • Opinionated towards DeepSeek API. I still don't know how to turn off their default web search tool from settings. It requires a DeepSeek API key and fails when models call it if a key is not provided. Even when I setup my own MCP for web search, models default to attempting to use the built-in tool and fail every time (then tries the mcp tool).
    • Not clear which parts of the harness can be changed on the fly and update immediately, and which parts require a full harness restart to apply.
    • Has the same issue as opencode: uses a hardcoded list of models per provider rather than fetching the model list from v1/models endpoint; leading some providers to show a stale list of models (such as qwen's token plan).
    • Functionality breaking bugs but that's expected

    If you've tried it, what's been your experience with deepseek harness?

    From a total noob: Are you using strix halo? in c/localllama@sh.itjust.works
    [–] 4 points 2 months ago (2 children)

    You don't need to worry about the split in strix halo, it runs a unified memory architecture so Linux will dynamically allocate VRAM on demand up to maximum possible (which would be around 120GB I think)

    I'm on an nvidia system so can't answer the rest of your questions other than to say that in my experience ollama gave me a much worse tokens per second speed compared to running the same model using llama.cpp, so if you get something slower than you expect it might be ollama caused.

    Good luck with your £250 system 😉

  • permalink
  • fedilink
  • source
    •
    12
    Hy4-Preview 770B-A49B preview frontier model from Tencent (programming.dev)
     

    Hy4 preview is a new-generation Mixture-of-Experts (MoE) flagship model developed by the Tencent Hy Team. The model comprises 770B total parameters, of which 49B are activated per token. Model supports 1M tokens context length.

    Native FP8 weights are published as well.

    Benchmarks:

    Known issues with the preview model: spends longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work.

    Open-weights frontier models will continue to drop until RAM prices improve 😅

    Qwen3.8-Flash-Next Weights Released (125B-A6B) in c/localllama@sh.itjust.works
    [–] 2 points 2 months ago*

    I've been testing this model since yesterday through the Qwen API in opencode and deepseek harness. I actually prefer it over GLM-5.3-Flash! it's super good at agentic coding and comfortably fast, now my goal is to be able to one day run it locally 😳 I think a Strix Halo with 128GB and a fast nvme should be able to run it well at near loseless quant

  • permalink
  • fedilink
  • source
    •
    Qwen3.8-Flash-Next Weights Released (125B-A6B) in c/localllama@sh.itjust.works
    GLM-5.3-Flash: Frontier Intelligence, Flash Cost 320b-a18b in c/localllama@sh.itjust.works
    [–] 4 points 2 months ago

    This is better than GLM-5.2 which just 2 months ago was my go to high tier model!

    Also in OpenRouter this model is listed to have a maximum context of 1.31M tokens! this is the first high-end model to my knowledge that supports more than 1M tokens context 🤯

  • permalink
  • fedilink
  • source
    •
    DeepSeek V4 Pro 0813 Weights Released! in c/localllama@sh.itjust.works
    Ornith 1.5 (Qwen finetunes for coding) released 9b, 35b-a3b, and 397b-a17b in c/localllama@sh.itjust.works
    [–] 3 points 2 months ago (1 child)

    Yeah I tried Kat Coder when it came out, I don't remember the exact issue I ran into but I identified a definite regression in my specific use cases compared to using vanilla. Ornith seems to be equivalent or better so far 👍

  • permalink
  • fedilink
  • source
  • parent
  • context
    • Ornith 1.5 (Qwen finetunes for coding) released 9b, 35b-a3b, and 397b-a17b in c/localllama@sh.itjust.works
    [–] 1 point 2 months ago (4 children)

    My previous experience with finetunes is bad..

    Today I tried Ornith-1.5-35B-A3B, and I'm quite positively surprised by it.. I compared it to both Qwen3.7-Flash and MiMo-V2.5 (via OpenRouter) in a few coding tasks and Ornith seems to be better than both! Thinks much more but the result definitely speaks for itself.

    Hoping for Qwen3.8-35B soon 🤞

  • permalink
  • fedilink
  • source
    •
    Low to midrange systems (8-32 GB) vs. free cloud tiers in c/localllama@sh.itjust.works
    [–] 3 points 2 months ago

    I'll add my vote to keeping the 32+8 system as that has the most ram.

    I also have an 8GB RTX 3070 Ti and 32GB RAM that I run Qwen3.6-35B-A3B perfectly on (30+ t/s) and looking forward to a possible Qwen3.8 35B update.

    Recently bought a used 8GB RTX 3070 for a bargain, planning to combine it with the 3070 Ti and hopefully run a lower quant Qwen3.8-27B which is crazy good for its size.

  • permalink
  • fedilink
  • source
    •
    Qwen3.8-27B Weights Released in c/localllama@sh.itjust.works
    DeepSeek V4 Pro 0813 Weights Released! in c/localllama@sh.itjust.works
    [–] [S] 7 points 2 months ago (8 children)

    free to use if you have the hardware. For this model because of its size, the main problem is the hardware availability/cost. But in general there are 3 ways to run an open weights model:

    • pay a provider like DeepSeek/OpenRouter per usage
    • buy hardware that can run it locally: actually not a bad investment for a business
    • rent hardware that can run it from a cloud provider, hardware can be rented dedicated or time-shared (so called serverless).
  • permalink
  • fedilink
  • source
  • parent
  • context
    • GLM-5.3 Imminent drop in c/localllama@sh.itjust.works
    view more: next ›