Open-model news

Micron's chief executive told investors on 30 September that demand for its memory will exceed supply in 2027 and again in 2028 — conditions much tighter than 2026. Micron makes the memory itself, and memory is the single biggest input to the price of every machine we quote — roughly three times its late-2025 level already.

The details show how firm the shortage is: three-quarters of Micron's 2027 output is already committed, most customer conversations now concern 2028, and the high-bandwidth memory AI datacentres use is in large part sold out for 2027 at prices well above 2026. Micron's rival Samsung says the same — high-bandwidth memory will take almost 30 per cent of memory manufacturing capacity next year, up from a fifth this year, leaving less for ordinary computers (Reuters).

For a business or a household planning a machine, the practical read: what AI has made scarce is the memory, not the computation; new factories are being built, but they ramp up gradually; and a machine bought for a job keeps doing that job. Plan on memory being the expensive part for years yet — that is the makers' own forecast, not a guess.

On 28 September Ollama, the one-command installer a small business reaches for when it runs AI on its own machine, added a new kind of model. A decision model does not write an answer: given a message, a ticket or a document it returns a choice, a yes-or-no, or a score — each with the probability behind it.

That is the shape software needs for the everyday jobs: sorting the inbox, labelling the tickets, routing the request. The first two models are Nimble, a 9B from Bespoke Labs built for fast typed classification, and Tev1 from Together AI, in 4B and 0.8B sizes; both run locally on your own machine.

Early days, honestly so — the models ship as experimental editions, and this is a new class, not a replacement for the general models this site runs. But triage and sorting are where a great deal of everyday automation starts, and that work just became small, local and fast.

An open model has arrived at the level of the best closed ones. MiMo-V2.6-Pro is a trillion-parameter mixture of experts — 1.02 trillion in all, 42 billion working per token — with a million-token window, and it reads text, images, video and audio.

The maker's own benchmarks put it close to Claude Opus 5 and ahead of GPT-5.6 Sol on most of the agentic tests it ran; Artificial Analysis, which tests independently, ranks it first of the 114 open models in its class, ahead of DeepSeek's V4.1-Flash. The licence is MIT — free for commercial use, no strings.

The honest catch is size: this one belongs in a machine room, and the same release ships a 309B version and a 9B distilled model that runs on a single 32 GB card. The gap between open and closed is now measured in hardware, not in what the model can do.

Alibaba's Damo Academy and Zhejiang University's First Affiliated Hospital published RADAR in Science on 17 September: one model that reads a contrast-enhanced abdominal CT scan and reports 146 imaging findings across 18 anatomical structures, trained on more than 400,000 examinations — and in the study's reader tests, its assistance raised the detection sensitivity of 26 radiologists by about a tenth. The paper's own conclusion is the headline: generalist AI can now match human experts in routine and complicated radiology alike.

The code is Apache 2.0 and the weights are downloadable; the licence on the weights is non-commercial, so treat this as the direction of travel — research and evaluation today, and a licence worth watching for a commercial successor.

For a household as much as for a hospital, the scan is the paper that must not go to a cloud API — and an open model can now read it at expert level on a machine its owner controls.

PrismML rebuilt Qwen3.8-27B — a model this site runs — in ternary weights at 1.76 bits per weight, and the result is a 27B-class model in a file of about six gigabytes, running on a laptop or a single graphics card. The maker's own benchmarks put it at 98% of the full-precision model's average — reasoning, coding, vision and tool calling intact — in a file more than nine times smaller; it still reads images, holds a 262k-token window, and ran at 143 tokens a second on an RTX 5090.

It is Apache 2.0, and it went to the top of Hugging Face's trending list: 1.9 million downloads in its first four days. One catch before you try it: the files need PrismML's own build of llama.cpp today, and stock runtimes cannot open them yet.

The floor for a serious model just dropped — a household or a small business that was told a 27B needed the big machine can run one on the machine it already has.

Through September NVIDIA published official NVFP4 builds of Qwen3.8-Flash-Next and GLM-5.3-Flash (2 September), Qwen3.8-27B (4 September), GLM-5.3 (14 September) and DeepSeek-V4.1-Flash (16 September).

A four-bit build is about a third of the memory the full-precision weights need — the difference between a datacentre card and the one in the machine you already own — and it is the configuration this site quotes.

Take the vendor's own build over a community conversion: the calibration, the recipe and the licence travel with it.

One of the two models this site standardises on is now downloadable and free for commercial use: 763.2B parameters on disk — a 552B backbone plus 196B of Engram memory, with about 15B of vision encoder and auxiliary tensors DeepSeek does not itemise — activating 8B per token on the way in and 16B on the way out, so it runs far lighter than its size suggests. The cache is the real change: 890 bytes per token globally, about a quarter of the memory the previous generation needed, which is what makes a one-million-token window affordable on owned hardware.

It reads images, which the model it replaces did not.

DeepSeek is retiring V4-Pro, routing its traffic here until V4.1-Pro launches — and no V4.1-Pro is released yet.

The hub where nearly every open weight is published is changing hands.

NVIDIA's commitments are on the record: the platform stays open, NVIDIA hardware will not be required to build on or deploy through it, and multi-cloud, multi-accelerator support continues.

Nothing changes for a business running models today — but the shelf everything sits on now belongs to the vendor selling the hardware, and those promises are what to hold it to.

125B on the label, 176B once the separate n-gram embedding table is counted, and 180B packaged with the 4B MTP head — the model card's own numbers, with only 6B working per token — so it is far faster than its size suggests, and it reads text, images and video. Native context is 262k; the 1M figure needs manual configuration.

Read the licence before planning on it: the Qwen Community Licence 1.0 requires a separate agreement from Qwen for anyone running a "Model as a Service or AI Work Assistant business", with no revenue or user threshold. Internal use is explicitly exempt.

Our everyday models are Apache 2.0 and unaffected.

Z.AI shipped GLM-5.3 Flash on the same date as Qwen's release.

Two open releases in one day is now ordinary.

The question worth asking of each is not which scores higher but which one fits a machine you can afford to keep running.

A 22-billion-parameter open model that generates picture and audio together rather than separately, six to twenty seconds, up to 4K.

Worth knowing about and worth being honest about: video is still the most expensive thing to run locally, it needs the larger machine, and for most small businesses it is not a job at all.

The runtime behind most single-machine setups was in the b10600s by 28 August, released as sequential build tags rather than versions.

Good news for support of new models.

Also a reason to pin a build and test before moving, rather than tracking the tip on a machine somebody depends on.

A 32 GB DDR5-6000 kit is around $392, against $110–140 in late 2025. Reported 64 GB prices differ by source — roughly $850 to $1,118 — and we quote the range rather than pick one.

The cause is memory makers moving wafer capacity to HBM for AI datacentres, where a gigabyte costs about three times the wafer of ordinary DDR5.

This changes the price of every machine we quote, and it may last into 2027.

The best picture quality per gigabyte you can self-host, and MIT means no revenue threshold and no restriction on selling what it makes.

Eight billion parameters, so it fits the everyday machine rather than needing the large one.

Apache 2.0, 4.1 billion parameters, about 3 GB of memory. Nine languages, and it clones a voice from three to five seconds of audio.

Human listeners preferred it to ElevenLabs Flash v2.5 in 68 per cent of blind comparisons.

A European lab, a permissive licence, and it runs on your own machine.

Apache 2.0, 80 billion parameters but only 3 billion working per token, so it runs far lighter than its size suggests.

A 262,000-token window, which is enough to hold a real codebase.

Around 70 per cent on SWE-bench — the practical choice if you want a coding agent that never sends your source anywhere.

Apache 2.0 on the code and the weights, which is rare here. It writes clearer sung lyrics than anything else measured, including the paid services, and places second overall in listening tests behind Suno.

The trade is speed: it generates at about real time, so a four-minute track takes about four minutes.

Around 4.8 GB has to stay in memory once the audio codec is counted.

Fast enough to change the job — a four-minute track with vocals in under ten seconds on a mid-range card, under 4 GB for the smaller version. We are not stocking it.

The code is MIT but the weight files carry no stated licence at all, which is not the same as permissive, and several guides report the project as Apache 2.0 when it is not.

Until that is fixed you cannot safely sell what it makes.

Used RTX 4090 cards are trading around $2,100, up from $1,200–1,400 earlier in the year: production ended in 2024 and demand keeps absorbing what is left.

If you were planning to save money with a used card, price a new one before you decide.