sounds like it’s time to take local models seriously

Samuel

@samuel.fm

Unsloth AI@unsloth.ai· 3 Apr

Gemma 4 E4B (4-bit) completed a full repo audit by executing Bash code and tool calls locally. Runs on just 6GB RAM.

7 8 155

Josh 4 Apr

with Claude’s shit limits lately, I’d love a local model

Steve Schmidt 3 Apr

Appears to be one of the few local models worthy of serious exploration and use.

Joel 4 Apr

Just had this discussion at an event in Amsterdam We got a bunch of unicorn CTOs together and had a full out discussion about it Consensus: LLM api agents are the cheapest they’ll ever be. The agents of the future will be more specialized instead of generic and then agents will coordinate

Tim 3 Apr

i was expecting to see TurboQuant mentioned but i guess this doesn't even include that yet, which is all the more impressive

2 1

Thoth 4 Apr

They have a 26b moe model competing with GLM-4.7 on the benchmarks. Which like. Lies, damned lies, and benchmarks, sure. But those are some impressive damned lies!

Daniel ❄️4 Apr

TurboQuant comes with some impact on tok/s despite taking less RAM, at least so far. Might be insurmountable just as a consequence of it being compression github.com/TheTom/turbo...

GitHub - TheTom/turboquant_plus

Contribute to TheTom/turboquant_plus development by creating an account on GitHub.

github.com

Nate Codes 4 Apr

Runs incredibly well with Ollama on Apple silicone with MLX now ollama run gemma4 It’s that easy

asura 4 Apr

It has been time to take local models seriously Now the challenge is parallel to getting people weaned off twitter - definitely rational people that certainly do not have pet sand are absolutely not obsessed with their big boy, but can't build local model workflows or pipelines when it's just .md

Thoth 4 Apr

Really hoping to get something useful at reasonable performance out of the 26b A4B MoE model they put out. Hoping we may have a model that is both genuinely useful and could plausibly run on a high-end PC.