nmfisher

Still there for me: "The full model weights will be released by July 27, 2026"

If you're on Bandcamp or Soundcloud it's usually because you want to support artists directly, I doubt many people are purely interested in getting free music rips.

Completely agree. Slow but smart models (Fable, Sol, GLM5.2 etc) are great, but they leave me with zero mental model of the code that's been written. Most of the time my mind wanders off and I go check social media or fire off a prompt for some other random project, it's a big productivity drain.

Working with models that are super fast, but slightly dumber (like mimo-v2.5-pro-ultraspeed) is amazing, I feel like I'm still the one that's actually making every decision.

Or he just finds it an incredible time-saving tool to help him do more maths.

I'm not a regular v2.5 user, so I don't really know. But given the TileRT team write-up says the entire network gets quantized to FP8 (and experts get quantized to FP4)[0], I'm assuming there's at least a modest drop in quality.

[0] https://www.tilert.ai/blog/breaking-1000-tps.html

Related, I was given access to mimo-v2.5-ultraspeed, which is amazing. This is now my expectation for speed, it’s fast enough for me to stay mentally engaged rather than getting distracted waiting for the agent to churn.

I was playing with Hy3 via openrouter yesterday (and I've also been using DS4 Flash/Pro as a daily driver since I cancelled my Anthropic sub a week ago).

I've found DS4 Flash to be very temperental (via Claude Code). The speed is great, but it often builds a completely wrong mental model and charges off down the wrong path. I find myself needing to rein it in regularly (and also compact the history, which undercuts the whole cache price advantage).

Hy3 isn't as fast, but so far it seems to stay on track much more reliably than DS4 Flash. It also doesn't seem to degrade as much with longer context. I'm not sure what the real pricing is, but I feel like it's a very competitive model.

As an aside, I also nabbed a 50m token pack for LongCat 2.0 to give it a whirl. Not free, but it's so cheap they're basically giving it away. Very impressed too - seems roughly on par with Hy3. Not frontier-level intelligence, but a dependable workhorse that can navigate a codebase well and can reliably execute what you tell it to do.

Maybe this is what Medusa Halo will turn out to be?

If you look at everything through their safe AGI mission it all makes sense.

Except for, you know, all the outside investors and the forthcoming IPO.

Not saying the results were bad - quite the opposite. But it was very slow (and if I was paying API rates, hideously expensive).

From my brief window of Fable usage, speed wasn't its strong point at all.

For actually building software, I'm starting to suspect a human with a dumber (but faster) model is going to get the job done quicker than Fable (and possibly even cheaper). Bug-finding and vulnerability detection is a different story.

I haven't tried their Hermes agent yet, because I only want a coding agent and I wasn't sure if theirs was suitable. Would you recommend it?

This is a blast from the past. Knoppix saved my life a few times, it was the easiest way to mount a drive with a broken partition table or something else went that haywire with a dual-boot system. It was also the safest option for doing something on a public computer without leaving a trace (though back then NIC drivers were always a bit finicky).

My first Knoppix CD may have actually come by way of the front cover of Linux Magazine.

Quick Google suggests Uber is up 66% on its IPO price, but the S&P index is up 85% over the same time period. I think Softbank also sold out around 2022, so the return (vs IPO price) would have been even lower. Didn't check for stock splits etc but I don't think Uber was a home-run for Softbank at all.

The investment round only closed in the last few weeks, it would have had zero influence on anything up to & including DeepSeek V4.

Whether that now changes, who knows. It appears the CEO will remain the largest investor/shareholder, though, so I'm hoping not much changes.

It makes no difference to me if a coding model has an opinion about Tiananmen Square, Americans bombing schoolgirls in Iran, how many genders there are, or anything else other than designing and writing code.

A coding model is a tool, as long as it follows its user's instructions for building software I don't really care what opinion it spits out about world history.

Yes, it is important to ensure that aren't hidden guardrails that are affecting its ability to perform its function. But the great thing about open weight models is that you can actually evaluate this rigorously, and retrain to remove any prejudices you don't like.

Until recently, DeepSeek were self-financed (it was a spin-out from a hedge fund). They just raised ~50million RMB (US$7bn), and according to media [0] (which admittedly can be unreliable), the lead investors were:

1) The CEO himself 2) Tencent 3) CALT (the battery company) 4) NetEase (internet/media company) 5) JD.com (ecommerce) 6) Chinese investment firms

What are they expecting in return? I'd say the same thing that all those investors in OpenAI and Anthropic are expecting - profit.

[0] https://finance.sina.com.cn/stock/vcpe/2026-06-11/doc-iniazi...

What you said is not "a matter of fact" because it's simply untrue.

These companies were not "created by" the Chinese government. Specifically, I'm talking about DeepSeek, Zhipu, MiMo (Xiaomi), Kimi (Moonshot), Qwen (Alibaba). "Subject to" certainly does not mean "created by", it just means that the government ultimately has the power to tell them what to do. The US government has the exact same power, hence why none of us has access to Fable at the moment, but you wouldn't say that OpenAI or Anthropic were "created by" the US government.

There is zero evidence that open-sourcing their models is part of some grand strategy from the Chinese government. In DeepSeek's case, I think it probably is a genuine commitment to open source, for the others I think it's probably just a convenient business decision to gain market share (though Zhipu is probably more aligned, given their academic lineage from Tsinghua).

At some point in the future, the Chinese government may decide it's not in their national interest for Chinese companies to open source their frontier AI models, and DeepSeek et al will be restricted from doing so. I'm well aware of that. But until that point in time, the rest of the world is unanimously better off with open-source Chinese models. We should put as much reliance on Chinese companies long-term as we do on American companies - zero.

None of those companies are created by the Chinese government. They're obviously subject to the Chinese government, whose whims may change at any given moment, but as we're seeing at the moment, so are the American companies.

And while I don't have a very positive view of the Chinese government, last I checked, they haven't been dropping bombs on innocent schoolchildren recently.

With current paradigms, yes. I'm hoping to see more focus on architectures that are more amenable to distributed training in the near future.

More like one bootlegger complaining that another bootlegger is copying their bootleg DVDs.

Facebook is the 21st century version of the tabloid newspaper and daytime Jerry Springer. Highly lucrative, but bottom of the barrel stuff.

Launching a prediction market is exactly the type of thing I would expect from them.

Wouldn’t you kind of expect to be fired if you release a project under your employers name that’s not even associated with them and hasn’t been cleared?

Not really, no. I'd expect a stern reprimand, but getting fired is extreme.

I'm not sure if Google is still an attractive place to work, but this incident certainly isn't helping tip the scales.

I doubt owners and banks actually believe this.

I worked for a commercial property company before, and yes this is exactly how it works, and yes it's just as stupid as it sounds.

Margin Call nailed it perfectly:

"It's just money. It's made up. Pieces of papers with pictures on it so we don't have to kill each other just to get something to eat".

It's not immediately clear, but this seems to be 250 tok/s on an M4 Max.

For comparison, the current agent swarm challenge on HF is at 508 tok/s on a A10G GPU:

https://huggingface.co/spaces/gemma-challenge/gemma-dashboar...

Half of the opposition to any regulation is the risk that it gets misused.

It did install, but then would hang when trying to configure the provider (specifically I was trying diffusiongemma via Nvidia). I gave up after a couple of attempts and moved to the next one.

Good timing, I was looking for alternatives earlier today. opencode didn't install properly and I wasn't a fan of oh-my-pi and nanocoder.

MiMo code (via my z.ai coding plan) is very pleasant so far, nice UI and seems to respond faster than Claude Code. It might be injecting much less cruft into the conversation.

I also got access to the mimo-2.5-pro ultraspeed model yesterday, which is really quite snappy. It does cost more than DeepSeek, though, so I'm not sure whether it's worth it yet. Definitely fast though.

I cancelled mine immediately too. Anyone who supports open models will sympathize.

It's shady because they were going to silently poison your outputs.

It's actually worse than it sounds initially, because Fable isn't actually omniscient when it comes to safety classification. Many people (myself included) had refusals or fallback to Opus 4.8 for seemingly compliant/innocuous requests.

Wouldn't you be pissed off if they decided to sabotage your project despite having done nothing wrong?