Apple's Mac Studio with 512GB of unified memory now has a price: a $15,500 upgrade over the 96GB base M5 Ultra, or $22,299 with 1TB storage (9to5Mac) - more than double the roughly $9,500 the Apple Angle commentary cited in July.
Apple's Mac Studio with 512GB of unified memory now has a price: a $15,500 upgrade over the 96GB base M5 Ultra, or $22,299 with 1TB storage (9to5Mac) - more than double the roughly $9,500 the Apple Angle commentary cited in July.
Open models keep getting smaller, cheaper and more specialised, and tools like Ollama make them easier to run on your own machine. Bigger models still cost more per token to serve, and the hardware needed to run the largest ones locally is getting pricier too. The shift away from paid APIs is real but not free.
Today's releases are mostly specialised open models (speech, document parsing) from large vendors rather than new general frontier models. They support the thesis that "good enough and free to download" now covers more everyday tasks, though the benchmark claims below are vendor-reported.
Nvidia released a free, open-weight transcription model covering up to 40 languages. TechCrunch reports it has 30B parameters, runs on a single chip (from Jetson AGX Thor to Blackwell datacenter GPUs), handles recordings up to an hour with up to eight speakers, and works on recordings only, not live audio. Nvidia claims it beats rivals on benchmarks including Hugging Face's Open ASR Leaderboard (vendor-reported); an open speech dataset, Granary v2, was released alongside.
TechCrunch · Open-weight model · October 28, 2026
Baidu released an MIT-licensed 3B-parameter OCR model that parses hundred-page documents in a single request. Its attention design keeps per-step decoding cost nearly flat, and Baidu reports a score of 77.1 on MarathonDocBench (vendor-reported; based on the feed summary, as the full article could not be opened).
MarkTechPost · Open-weight model · October 26-27, 2026
One item today bears directly on the token-price argument: a senior OpenAI figure restates that bigger models cost more per token. This is a general statement, not new pricing data, so it is evidence for the thesis only in a weak sense.
Simon Willison quotes Ted Sanders explaining OpenAI's model naming. Sanders says much of the variation in LLM performance is explained by model size, and that bigger models will be more reliable, cost more per token, and have better recall of esoteric facts. No prices or parameter counts are given, so treat it as framing rather than a data point. It does underline why a cheaper open model can be attractive when a smaller one is good enough.
Simon Willison's Weblog · Analysis · October 28, 2026
Apple's top memory option and a lighter-weight local-model runtime update both landed today. Together they show the hardware ceiling for local models rising in capability and in price.
9to5Mac reports that the 512GB option is only available with the top M5 Ultra (36-core CPU, 80-core GPU) and is a $15,500 upgrade over the 96GB base configuration. The total is $22,299 with 1TB of storage, up to $29,799 with 16TB. 9to5Mac says orders arrive in late November to early December; I have not verified availability independently.
9to5Mac · Hardware · October 28, 2026
LFM3 models, including the text-to-speech in lfm3-audio and lfm3-omni, previously Apple-silicon only in Ollama, now run on Linux and Windows too. This is a follow-up to the LFM3 release covered earlier this week and widens the pool of machines that can run these small models locally.
Ollama releases (GitHub) · Follow-up · October 29, 2026
UC Berkeley's open-source LEANN stores a pruned graph and recomputes embeddings at query time. The team reports it matches HNSW accuracy at 90% recall using 4 GB instead of 188 GB, at the cost of seconds-level search (researcher-reported; based on the feed summary, as the article could not be opened). Smaller memory footprints make private, local retrieval setups more practical.
MarkTechPost · Research · October 27, 2026
Some commentators argue Apple, not Nvidia, is best placed for the on-device and on-prem shift. The case in the linked piece: a single Mac Studio with unified memory can hold very large open-weight models that would otherwise need five or six Nvidia RTX Pro 6000 cards (roughly $60,000-$75,000, with around ten times the power draw), and as capable open-weight models keep shipping, the bottleneck moves from who has the best model to who has affordable hardware to run it locally. The article, dated July 17, cites a 512GB Mac Studio at about $9,500; today's 9to5Mac report puts the 512GB M5 Ultra option at $22,299 with 1TB, so that price point no longer matches what Apple is listing. Counterpoint: Mac Studios suit single-user, high-memory inference, while datacenter GPU clusters still serve many users at once.
Limited Edition Jonathan (Substack) · Perspective
Open-weight models are still being released at a steady pace and the software to run them locally keeps widening, which supports the shift away from paid APIs. But the cost of high-memory local hardware just rose sharply, which complicates the "cheap on-device" half of the thesis.
Keynotes, masterclasses, panels and board-room sessions.