1broseidon

holy crap, thank you. Its actually running at 54 tok/s which is incredible for local inference.

I’ll check it out. I tried the unsloth version but it couldn’t handle anything larger than an 8k context window unless I used a 2 or 3 bit quant

I “upgraded” from an Ubuntu desktop with ryzen 9 and a 3090. I still have the box but it runs like a sauna.

My new set up is a Mac Studio M5 Mac 128gb. Surprisingly it’s still hard to find decent models to fit on 128gb, but Qwen3.8 27b runs FAST so it’s really great as a personal agent that isn’t quantized to heck

ah there it is, it already exists!

For sure, and "content policy" is broad enough that it could be at their will for what they decide is "risky" - but imo, it just invites more shady behavior via decentralized protocols we already know exist. Torrents, NZB, etc. The risk now moves to the consumer who might download something that exfiltrates all their data.

Doubt it. Even if HF drops them, it will not be hard for someone to move it to something like a decentralized torrent site dedicated to open source models.

Banning them would just invite competition which is why I think HF is holding out as long as they legally can.

I get that Claude wants to know the user and have all the context, but this seems like a bad decision by the LLM that should have been caught in Auto Approve mode.

Primarily training data. A lot of times it’s a hedge to show two things and let the user decide which fits their framing best

Most of the large models collected data from marketing speak, opinion posts, etc (the whole internet basically). I feel like it’s a common sales pitch I’ve heard many times over - now it’s just compounded with everyone having access to the same pattern generator clanker

even that was relaxed by Anthropic and they stopped chasing token prices for Agent SDK. I routinely use Agent SDK with my subscription and it doesn't come out of a separate budget. Their ToS vaguely says just don't use your personal subscription for commercial resale purposes.

This is just to make me feel poor for "only" having an M5 Max 128GB model. Which is insane compute for a home computer, but now I already feel behind one month into it.

Both- depending on the work. CLI for anything local and deterministic: code tools, builds, git, file ops. The agent already knows how to exec commands and parse output, and when something breaks you get stderr and an exit code. MCP for discovery and remote services where the agent needs to figure out what's available.

They solve different problems, but imo MCP has a speed overhead problem that cli tools don’t usually have - so my preference is CLI.

100% - you are very right. 4.6 is amazing for orchestration. I even built some tools around agent to agent contracting.

I use 4.6 as the brain and then handoff to a more rigid llm like GPT 5.2 or Opus 4.5

I think it’s kinda one of those YMMV metrics. I work for a large company around 80-100k employees and they are ranked on this list. The business unit I’m in is amazing and I truly feel that it’s a great place to work. But ask someone in a dying business unit how they feel and they’ll say they’re overworked and understaffed.

I have found this to be true too and I thought I was the only one. Everyone is praising 4.6 and while it’s great at agentic and tool use, it does not follow instructions as cleanly as 4.5 - I also feel like 4.5 was just way more efficient too

I think the 'AI productivity gap' is mostly a state management problem. Even with great models, you burn so much time just manually syncing context between different agents or chat sessions.

Until the handoff tax is lower than the cost of just doing it yourself, the ROI isn't going to be there for most engineering workflows.