The AI models worth knowing, checked by hand: the licence, whether you can download the weights, the price with the date it held, and who ran the benchmarks.
“Runs on” means the model’s weight file, a 4-bit build where one exists, takes no more than 75% of the machine’s memory, leaving the rest for context, the runtime and the operating system.
DeepSeek's weekday peak price (01:00–04:00 and 06:00–10:00 UTC); every other hour is half. The median across 25 providers is $0.29 in, $1.14 out. Price source.
No independent benchmark in the registry yet — vendor tables are deliberately not reproduced here.
What changed, and does it matter
It replaced V4-Flash outright: ask DeepSeek for the old model and you now get this one, at the Flash price. What changed is how little it holds in memory while it reads, about 890 bytes per token by DeepSeek’s figure, a quarter of V4-Flash. That is why a million-token context is cheap for DeepSeek to serve. It is not yours to run yet. The release is 510 GB, and nobody had published a 4-bit build when we checked.
Tied with Kimi K3 for the best blind-vote score here, at a thirtieth of Kimi’s output price. It is the first GLM-5 model that reads images natively, and it mixes two kinds of attention to keep long documents cheap. More providers sell it than any other model here. Rent it, though: the smallest 4-bit file is 200 GB.
Median of 16 providers on OpenRouter; the range is $0.10–$0.45 in, $1.80–$3.20 out. The consumer-hardware story is the point of this model. Price source.
The model that fits a 24 GB graphics card when the arithmetic says it shouldn’t. It has 64 layers, but only 16 use full attention; the rest use a cheaper kind (Gated DeltaNet) whose memory does not grow with the conversation. So the memory it needs to hold a long conversation is a quarter of what 64 layers suggests. That is the difference between "cannot possibly fit in 24GB" and "fits with room to spare", and it is the subject of the sizing tutorial this entry backs.
DeepSeek's weekday peak price, which is also the median across 24 providers. Off-peak is half ($0.66 in, $1.98 out), which is what the whole day cost in August. Price source.
An MIT-licensed model you can download, shipped with a harness (the software around the model that lets it use tools) whose plugins you can swap. Watch the price, because it keeps moving. It went up three days after launch, and DeepSeek has since split the day in two: the August price now holds only off-peak, and weekday peak hours cost double.
Built for long documents. It mixes Mamba-2 layers, whose memory stays flat however long the document gets, with mixture-of-experts layers (only a slice of the model works on each token) and just 6 attention layers out of 52. NVIDIA ships several ways to speed up generation. The 4-bit file is 25.3 GB: just over our line for a 32 GB machine, easily inside 64. The catch is quality. Its blind-vote score sits about 100 points under the Gemma 4 and Qwen models of the same size.
The dearest model here, at nearly four times the next one’s output price. Moonshot calls it the first open model in the 3-trillion class, and it reads text, images and video in one model. It ties GLM-5.3-Flash for the best blind-vote score here, measured at its most deliberate setting. The weights are open under Moonshot’s own licence, not a standard one, and the 4-bit file is 1.5 TB. Open weights, but you will rent them.
The score you can trust most here: 48,540 blind votes, more than any other model in this registry, so it is the least likely to move. It reads images natively and holds a million tokens of context. The licence is MiniMax’s own, so read it before you build on it. At 264 GB for the 4-bit file, it runs on a server, not a desk.
Median of 9 providers on OpenRouter; the range is $0.05–$0.25 in, $0.70–$1.80 out. Price source.
No independent benchmark in the registry yet — vendor tables are deliberately not reproduced here.
What changed, and does it matter
The biggest model here that still fits a 32 GB machine by our rule: 22.1 GB of 4-bit weights. Only 3B of its 35B parameters work on each token, so each token is little work for a model this size. People are running it: NVIDIA’s and Qwen’s own compressed builds both sit in Hugging Face’s 50 most-downloaded models. It has no blind-vote score yet, so we cannot rank it against Qwen3.8-27B.
On a 32 GB laptop, test this side by side with Gemma 4 31B. It scores 8 points lower in the blind vote, but it works only 3.8B parameters per token, about an eighth of the 31B’s load, so on the same machine it has far less to compute. The 4-bit file is 17 GB.
The best-scoring model here that fits a 32 GB machine, from an 18.3 GB 4-bit file. It is also the first Gemma under a standard open licence, Apache 2.0; earlier Gemmas came under Google’s own terms. That matters if you build a product on it. It reads images as well as text.
OpenAI’s first open weights since GPT-2, shipped already squeezed to 4 bits, so the file is about 63 GB whichever build you pick. OpenAI sized it for one 80 GB graphics card. By our rule it needs a 96 GB machine, not a 64 GB one. It is a year old, and it shows: its blind-vote score sits roughly 80 to 100 points under this year’s open models.
The one to try on a 16 GB machine. At 11.6 GB it is the smallest file here, and the only model that fits 16 GB by our rule. It is also the cheapest to rent. You pay in quality: its blind-vote score is the lowest here, about 150 points under Gemma 4 31B.
