Daily AI Fresh News · Oct 30 Evening Edition: DeepSeek's Double Open-Source and the Free Inference Retreat

Lots of new moves in the AI world today, and a bunch of them are pretty juicy.

Here's what we're covering today. First, those GPT-6.2 leak rumors. Second, Ollama's free tier switching to paid. Third, Waymo teaming up with Toyota to build cars. Fourth, the document parsing model OXR. Fifth, Qwen breaking when run locally. And finally, the big one: DeepSeek^Ad's open-source inference engine and the phone version.

Yeah, all of those are definitely worth talking about.

Alright, let's run through the news quickly first, then we'll dig into the DeepSeek stuff.

First one — I just saw a video on YouTube, and the title honestly broke my brain. It says GPT-6.2 leaked, plus some kind of debut for Opus 6 and Gemini 4 Carbon.

Titles like that are pure clickbait. But that's not the interesting part — the interesting part is it also squeezes in Claude Fable 5.5 and Qwen 4.

So how credible do you think it is?

I lean toward it being mostly guesswork. Think about it — Fable 5 just came out, and 5.5 already leaked? The timing doesn't add up.

Fair. Classic clickbait. But people always shout this stuff before a big release — good enough as background noise.

Second one's pretty concrete — someone on Reddit posted the official email. Ollama's Free plan is automatically switching to pay-as-you-go on November 9th, with prepaid credits.

The signal is clear: the free inference window is closing. GPUs sitting idle still cost money.

Right, and the email lists them out — gemma4:31b, the big and small gpt-oss, plus several nemotron-3 versions. All going paid.

The article's three options are pretty practical: top up a little, move back to local, or switch cloud providers. If you're not running big models, going local is the easiest.

Okay, third one's a change of pace — Waymo and Toyota are joining hands. Toyota supplies the bZ4X EVs, Waymo drops in its Driver system.

I really like that split. Toyota handles manufacturing and local compliance, Waymo focuses on software — way more scalable than grinding out your own fleet.

They'll test in Japan first next year, then the UK. But fleet size and whether they'll carry passengers — nothing's been said.

Yeah, details are still under wraps. But this path shows the top players are shifting to a hybrid model — keep the core algorithms in-house, buy the cars and resources outside.

Fourth one's more technical — the SoMarkAI team on Reddit open-sourced a document parsing model called OXR, 1.3B parameters.

It tackles the big headache of PDF reading order. Think about it — get the order wrong on a multi-column layout, and the whole downstream NLP pipeline collapses.

How does it do it?

It breaks layout analysis into swappable modules, then predicts elements and order together in a single forward pass — no two-stage error stacking.

Sounds clever. But there's a hard barrier — running locally needs 24GB of VRAM, and multi-column layouts are still a weak spot.

True, but it ranks first on three benchmarks, and the weights are open-sourced. The document RAG crowd is probably thrilled.

Last quick one — someone on Reddit complained that Qwen 3.5 9B on Ollama dumps the entire thinking process, then the final answer never shows up — just a cursor.

That's so classic. The hardware can run it, but the model template mishandles the end-of-thinking token, or num_predict truncates the generation.

So the troubleshooting order is: check the Modelfile template first, then stop tokens and length parameters?

Exactly, and compare against non-thinking mode too. Thinking models bring new headaches to the local toolchain — meeting the hardware bar is just step one.

Alright, that's the news. Now let's dig into a topic that's been blowing up on Bilibili lately — DeepSeek.

There's a video from a French blogger with a really catchy title — says America's still burning cash while Chinese AI is printing money. The core is DeepSeek's open-source inference engine, claiming a single GPU pays for itself in two months.

I looked at that math. Let me put the conclusion up front: the single-GPU two-month payback claim is directionally right, but take the numbers with a discount.

Why a discount? How did they calculate it?

It compares inference throughput against electricity and rental costs. The problem is it assumes full load — and don't underestimate that assumption, because in reality it's hard to stay maxed out.

So it's treating the ideal case as the norm. But there's a reason it blew up — open-sourcing the inference engine is a big deal on its own.

Right, the key is open source. What does open-sourcing the inference engine mean? It means the cost black box of inference is cracked open — smaller players can build it themselves.

There's debate on Reddit about whether this commoditizes inference. Others pour cold water, saying nobody's counting the hidden costs of deployment and tuning.

The programmers on HN are even more blunt — they say payback period depends on your QPS, not on some blogger's screenshot.

But isn't that contradictory? One side says huge profits, the other says high hidden costs.

Not contradictory. Open source drives down software costs, but hardware and ops are still hard expenses. You're saving on license fees, not electricity.

Got it — it's swapping money paid to closed-source vendors for buying your own GPUs and tuning them yourself.

Exactly, so that huge-profit math is more of a传播 talking point. What's really valuable is that open-sourcing the inference engine makes the cost structure transparent.

Now the second one — DSH-GO, the mobile version of DeepSeek Harness, with the slogan 'AI officially takes over your phone.'

This one actually interests me more. Running a harness on mobile — the barrier is VRAM and power. The fact it runs at all shows the on-device path is moving forward.

But it still feels like a demo — actually taking over the phone is a long way off, right?

Far off, but the direction is interesting. Remember that data showing Android on-device AI was twice as fast as cloud — the economics of on-device inference do work out.

So my read is: on one hand, open-sourcing the inference engine drives down cloud costs; on the other, pushing onto phones — squeezing from both ends.

To sum up, the core of this DeepSeek wave isn't huge profits — it's turning inference from a black box into something you can build yourself. Cost transparency is the real impact.

Alright, that's all for today. We'll keep an eye on both DeepSeek stories. See you tomorrow!