Skip to content

MegaCapybara 1.36

Choose a tag to compare

View all tags
@perkel666 perkel666 released this 20 Oct 21:12
· 5 commits to main since this release
v1.36
f107124

What's new

A new model: Qwen3.8 Flash-Next

MegaCapybara now runs Qwen3.8 Flash-Next, a 177-billion-parameter mixture-of-experts model, on one RTX 5090. It is built on turboderp's EXL3 quantization and runs on MegaCapybara's own kernels.

The model's experts don't all fit in the card's 32 GB of memory. The most used ones stay on the GPU and the rest are read from RAM. While you work, the engine moves the experts your work needs most into VRAM.

  • Two sizes. Tiny stores the experts at 2 bits and is faster, with more room for context. Small stores them at 3 bits and gives answers closer to the original model.

  • Uncensored versions of both sizes. The launcher downloads all four from perkel/Qwen3.8-Flash-Next-MC and perkel/Qwen3.8-Flash-Next-Uncensored-MC.

  • Two ways to write faster: DFlash2 (the default) and MTP. DFlash2 uses a separate drafter that the launcher downloads with the model. MTP uses a layer built into the model. Both give exactly the model's own answers. Speed with Tiny, in tokens per second:

    No speculation MTP DFlash2 (default)
    Coding, one agent 299 440 498
    Prose, one agent 289 300 318
    Coding, four agents at once 678 789
    Coding, eight agents at once 779 993

  • Images. Flash-Next can read images. Its vision encoder stays in RAM, so it takes no GPU memory away from the experts.

  • Long context. Conversations can be up to a million tokens long. You can store the context at 8, 6 or 4 bits per value to fit more of it.

Suggested: 64 GB of RAM.

All models

  • Agents that don't use a drafter now share their work. Before, each one ran in turn. Now they are processed together, so eight agents on the 27B Medium model reach 622 tokens per second instead of 92.
  • Images are processed faster. A one-megapixel image now takes 50 milliseconds instead of 221.

Fixed

  • A short opening that many requests share is now kept in memory. Before, it could be pushed out by longer ones and read again every time.
  • Conversations that have just answered are no longer the first to be removed when the RAM cache is full.
  • Emoji and other characters that span two tokens came out as � in tool-call arguments, which could break an agent's file edits.
  • When the loop guard stops an answer, the OpenAI API now reports finish_reason "length". It used to report "repetition", which some clients such as OpenCode rejected.
  • The server could fail with "save KV: out of memory" and stop answering when many experts or conversations were held in RAM. It now stays within the memory that Windows lets the GPU use.
  • Requests that ask for no thinking are now honoured in every common form.
  • Requests are no longer limited in how many images they can contain.

Launcher

  • The download panel groups models by family, Qwen3.8 27B and Qwen3.8 Flash-Next. One Download button gets a model with everything it needs.
  • The monitor shows a queued request and the conversation it continues as a single entry.

The "Source code" files GitHub attaches to this release hold only the README and the docs. The engine's source code will be released soon.

Full changelog · How to run it

Assets 4
Loading