nullptr 🐱🍩

scumfuck flowergirl // i know a little bit about computers // priv: @Zbuildstd

[nullptr 🐱🍩 @notnullptr]

@kalomaze i guess we’ll see how it holds up against the competition for transcription .. there’s no real niche here imo but we will see

[nullptr 🐱🍩 @notnullptr]

i’m glad google are still releasing open models. i just wish they were good models

[Google Gemma @googlegemma]

Meet Gemma 4 12B!

A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.

Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

[nullptr 🐱🍩 @notnullptr]

@kalomaze yeah but in a 12b dense model?? that’s a lot of weights dedicated to just the language side of things if you just want transcription

[nullptr 🐱🍩 @notnullptr]

@kalomaze the benches will be disappointing regardless. i think the selling point is the audio multimodality? but i don't see a use-case

[nullptr 🐱🍩 @notnullptr]

@f4micom the price is gonna be disappointing

[nullptr 🐱🍩 @notnullptr]

@forloopcodes parallel sub-agents can probably be made useful with enough guarding around file writes but people who literally have multiple harnesses open at once are stupid

[VERT.sh @VERT_sh]

You're welcome🙂‍↕️

[nullptr 🐱🍩 @notnullptr]

@analogalok i tested the full precision BF16 weights with F16 kv with stock llama.cpp inference settings @ 128k context (essentially a production deployment) under the pi harness. it struggles deeply with tool calls (outputting json sometimes?) and will often fail to close its own thinking

[nullptr 🐱🍩 @notnullptr]

@lunat1cpr1ncess doesn't demonstrate confidence in their own technology

[nullptr 🐱🍩 @notnullptr]

@Malix_Labs @giteaio @codeberg_org this is the sole reason i chose to use forgejo over gitea lol

[nullptr 🐱🍩 @notnullptr]

@spinelessaisha he bought a gpu cluster and hosts models himself .. i dont see the issue

[nullptr 🐱🍩 @notnullptr]

@arm64le the models have executive dysfunction, if you give a complex problem they're like "implementing the middleware... considering the architecture..." in the CoT it's so bizarre. and they output that long before any tool calls. very hesitant models

[nullptr 🐱🍩 @notnullptr]

@arm64le @evadotpink hiiii haha

[nullptr 🐱🍩 @notnullptr]

@arm64le that being said though you should give qwen 3.6 27b a try at some point . really cool little model, incredibly capable if you don't let it stop until an independent verifier makes sure its done a good job (ie i use a pi extension for /goal)

[nullptr 🐱🍩 @notnullptr]

@arm64le if i ever get my hands on 512gb of video memory you best bet i am running deepseek v4 at home

[nullptr 🐱🍩 @notnullptr]

@arm64le i trust the shitty little q5_k_m quant of qwen 3.6 27b with my LIFE

[nullptr 🐱🍩 @notnullptr]

@arm64le the coolest people give the chinese models full control of their computer

[nullptr 🐱🍩 @notnullptr]

@gf_256 LMAOOOOO

[nullptr 🐱🍩 @notnullptr]

@sarthak2143 my goat