A long AI agent review on X flags tasks marked "done" before they are. I'd underline that twice. One automation of mine looked finished. A second AI went through it and found 25 weak spots, a few of them real holes. So anything important now gets a second model trying to break it first.

For a few AI questions a week, the free version does the job. A $200 plan works out to roughly $10 per working day. Simple math. If it saves me more than that every day through real work, it's cheap. If not, it's an expensive hobby.

New to local AI with a 16GB card? Start smaller than you think. Qwen3.5 9B at Q4_K_M weighs 5.7GB and leaves well over half the card for the conversation. A 27B at 4-bit won't fit. It spills into system RAM, and then people decide local AI is slow.

A new CUDA engine runs Qwen3.8-27B at about twice llama.cpp's speed on one RTX 3090, in its author's own tests (spotted via QwenAIFans on X). Got a 3090? I'd try it first. 16 GB cards are out for now: the 4-bit file it's built around is 17.6 GB on its own.

What should the Macs run first? I've ordered 2 Mac Studios, 512 GB each. A fictional inbox, a JSON check, or a small coding task? Pick one concrete job I can repeat and measure. Delivery window: 17–24 November.

A text-to-speech model you can run on a laptop CPU, no GPU needed. Underdog Woof is one 40.6MB file. Their own test says it makes fewer word errors than models 1,000x its size. Big claim for a small file, so I'd judge it on my own text first. Personal use is free, paid products need a license.

Ollama v0.40.15 lets recurrent and vision GGUF models serve more than one request at a time. Agents don't take turns. This is the part I'd test first. The catch, from Ollama's own FAQ: the default is still one request per model. Four agents at 8K each need about the context of one 32K chat.

Inkling 2 Small is due to get its full weights next week under Apache 2.0, and the benchmark table is the last thing I'll read. First comes the real file size. 17B active per token sounds light, but every byte of it still has to load into memory.

Renting an open model by the API doesn't mean you get the model you think. Artificial Analysis checked 100+ hosted endpoints against its own run of the official weights, and nearly 1 in 4 scored below it. Same name, less model. On my box nothing changes unless I change it.

A model that says "I don't know" is worth more to me than a higher score. In Artificial Analysis' new test, Nanbeige4.3-33B declined 71% of knowledge questions and made things up less often than Qwen3.8 27B. I'd take that trade. The Q4 file is 20.8GB though, too big for my RTX 4080 😅

I like model changes I can actually read. Moonshot pushed one to Kimi K3.1 Pro today. A single public line on Hugging Face drops "medium" from the thinking levels, and the commit says why: it was invalid. Closed models change too. You rarely get to see the line.

Open coding models are getting small enough to own. Ripple-65B Preview came out yesterday under Apache 2.0, and the full-precision files are 130GB. One big-memory box, no rack needed. Ricursive's own card has it just ahead of a 501B model on TerminalBench 2.1. Real work decides.

Max thinking effort is mostly a waiting button. Back to my blind test on real business tasks: max gave Opus 5.5 half a point out of 200 over xhigh and roughly tripled the wait. GPT-6.1 Sol actually lost half a point on max. Not worth it for me. Those minutes go to the next task.

Mistral Large 4 is now on Hugging Face Inference Endpoints for $10 an hour, per ngxson on X. Renting a model that size by the hour is how I'd handle the jobs my own box can't. Just set a reminder. One forgotten weekend is $480 😂

SGLang shared day-0 serving numbers for Kimi K3.1 Pro on X, and they're datacenter numbers. At home the model hits a different wall. Under 4% of its weights work on any one token. All 2.8T parameters still need to fit in memory to run it at a usable speed. Most home setups stop there.

I don't call a model open until I can download it. Hy4 launched today, "firmly among the leading open models" in Tencent's words. Download link? Not yet. The only Hy4 weights on Hugging Face are still August's preview, so for now it's an API to me. Hope the final files follow soon.

People are typing "why local AI" into Google, and I love that question. US searches for it are up 1,286% in three months, from a small base. "what local ai can do" climbs even faster. I started out of curiosity. Now I just don't want one provider's limits deciding how much I get done.

Saw the "Opus 5.5 got downgraded" take, so I tested it blind on 20 hard real-work tasks. 7 setups, 2 runs each, 2 blind graders. Opus 5.5 came out on top at 197 of 200. Max effort bought half a point for 3x the time. Caveat: all 7 scored above 189.

For agents, a model that just picks from fixed answers is often more useful than one that chats. Sarvam's new Prism scores each option, like which team gets a ticket. Open weights, Apache 2.0, 44 ms median on one B200. The catch: about 120GB on a single FP8 GPU. Home setups wait.

Niklas Lenz on X put a Mac Studio against two DGX Sparks, each running a local model in a live strategy game he's building, nobody steering. I'd watch a whole league of this. Memory is why I picked Apple: two Sparks add up to 256GB, one 512GB Studio holds double. Mine should land later this month.

Renting AI compute for years and owning nothing at the end is a deal I'd never take. An H100 on Lambda's on-demand list starts at $3.29 an hour. Nonstop, that's about $28,800 a year, and the card goes back the day you stop paying. Cloud for hard jobs, sure. The everyday box should be mine.

Image-to-3D runs on normal home hardware now, and the weights are open. Speridlabs' Idol turns one photo into a 3D object and peaks at 7.2GB of VRAM by their own numbers. Plenty left on my 4080. Since last night it runs natively on Macs too, through mlx-vlm.

A cheap API price is only cheap if you check the clock. DeepSeek's Flash API costs twice as much in weekday peak hours, and on Central European time the longer window runs 7 to 11 in the morning, right when the workday starts here. My GPU at home? No surge pricing 😂

This is the kind of local AI progress I want more of: big open models on an ordinary gaming PC. tandem.cpp patches llama.cpp to keep a MoE model's busiest experts on the GPU. On the author's own RTX 4090, a model over three times the card's memory went from 29.8 to 97.3 tok/s.

On X, Deep Principle's Hydrogen-1.0 is called proof that open weights now beat frontier labs at science. The benchmark isn't the win for me. The lab trained open GLM-5.3 weights on its own lab records and owns the result, which no closed fine-tune gives you. Downloading it is another story: 1.5TB.

I wouldn't wait for local AI hardware to get cheaper. NVIDIA's new DGX Spark costs $1,000 more than the first one's old price, with half the memory. NVIDIA named memory supply as the reason when it raised that price in February.

Passing tests only count if they run against the real thing. My AI's tests had been talking to a mock the whole time, and the mock was built differently from the real API. The first real call was rejected on the spot. 67 green, first run red.

Apple charges more for an extra 256GB of memory than for a whole second Mac Studio. $11,500 for the step from 256GB to 512GB on the M5 Ultra, $10,799 for another complete 256GB box. The catch is running one model across two machines, and that's what I'll measure when mine show up.

TeksEdge on X floats a Qwen4 at 27B. That size would excite me more than any giant model, because people can actually run it at home. Today's Qwen3.8-27B is a 16.46 GB file at Q4. Nothing from Qwen itself yet though, so for now it's a rumor.

Two 512GB Mac Studios on order. Speed alone won't tell me much. Ivan Fioravanti's 8-user M3 Ultra test on X hit 191 tok/s, yet the 8th user waited about 9 minutes for a first word (he suspects the engine). Day one: fit, first word, tok/s, wall watts, euros per 1M tokens. What job goes first?