Voltropy unveils Vast-10M, a frontier model with 10 million tokens of context

Orange County lab Voltropy said Tuesday it launched Vast-10M, a three-size model family with a native ten-million-token context window, and opened early access on voltropy.com.

Frontier chat models have been stuck near a million tokens for years. If a 10 million-token native window holds up outside the lab’s own chart, whole codebases, year-long journals, and S&P earnings piles fit in one pass instead of a retrieval scavenger hunt.

On Tuesday, 29 September 2026, Voltropy unveiled Vast-10M. The company blog, titled “Introducing Vast-10M” and dated September 29, 2026, calls it the first frontier large language model with ten million tokens of native context. A large language model is the kind of system that reads and writes text. A token is a small piece of that text the model counts. Native context is how much text the model can hold in one pass, inside the model itself, rather than text a separate search step fetches and pastes in. The blog does not print an hour. PR Newswire carried the same launch, datelined Irvine, California, and stamps it Sep 29, 2026, 06:00 ET, which is 6:00 a.m. Eastern. The about box calls Voltropy an AI research laboratory based in Orange County. Irvine is in that county. Those lines are Voltropy’s.

The family ships in three sizes. Vast-10M-Flash is based on DeepSeek V4.0 Flash, which the wire writes as DeepSeek V4 Flash. Vast-10M-Medium is based on GLM-5.2, which the wire writes as GLM 5.2. The blog calls Medium the mid-size. Vast-10M-Pro is based on DeepSeek V4.0 Pro, which the wire writes as DeepSeek V4 Pro. Those bases are Voltropy’s.

What the three share is an algorithm Voltropy names Voltropy Scalable Attention, shortened to VSA. Attention, in a transformer, is the step that decides which earlier text matters for the next word. A transformer is the usual design behind today’s chat models. Voltropy says VSA widens that window and improves scores at shorter lengths, instead of trading intelligence for a bigger window. The blog says earlier approaches either made attention cheaper and less accurate, or left transformers for simpler designs, and that both bought more context by giving up intelligence. Those lines are Voltropy’s.

The scores Voltropy prints are on a test it calls BEAM. The blog spells that out as Beyond a Million Tokens. The wire calls BEAM an industry-standard benchmark. On the one-million-token tier, Voltropy says Vast-10M-Flash beats Anthropic’s Fable 5.1 and reaches parity with OpenAI’s GPT-6 Astra. The same post’s chart line says Flash beats Fable 5.1 by 4.70 points and reaches 97.95 percent of Astra’s score. A callout on the page says the Flash score is 10.68 percent higher than Fable 5.1 at one million tokens. The window, Voltropy says, is about 10 times Fable’s and 9.5 times Astra’s. Those comparisons are Voltropy’s. The pages do not print a rerun by an outside lab.

At ten million tokens, Voltropy says Vast-10M-Flash scores 40.20 on BEAM. It says that is above its DeepSeek base model’s score at one million tokens, 39.69, and that 40.20 is 101 percent of that base score at ten times the context. It also says Flash keeps 82.49 percent of its own one-million-token BEAM score when the window grows to ten million. An order of magnitude, the phrase on the page, means about ten times. The blog calls 40.20 the highest score any raw model has reached on that test, and higher than the scores it says are reported for most retrieval systems. A retrieval system searches a pile and pastes pieces into the model. Those figures are Voltropy’s chart.

What Voltropy says fits in ten million tokens: the entire U.S. tax code, earnings calls for every company in the S&P 500, the transcript of a multi-month trial, a full year of the New England Journal of Medicine, the TypeScript compiler, the SQLite database, and the React web framework. Clint Ehrlich, chief executive and co-founder, said frontier models have been stuck around a million tokens for almost three years, and that a person can paste in the tax code or those earnings transcripts. Ted Blackman, chief technology officer and co-founder, said earlier ways of stretching context made models less intelligent, and that VSA does the opposite at shorter lengths, so Vast-10M can compete even on tasks that already fit inside other models’ windows. Those quotations are theirs, on the wire.

Voltropy also says this is not a claim that Vast-10M is better than Fable or Astra at everything. The blog says those models have more raw intelligence on the hardest tasks, such as frontier mathematics. It says Vast-10M can do better on consumer and business work that needs both precise recall and reasoning. The same post argues that a large enough window is a path to a model that keeps learning after it ships, without retraining its weights, and that Vast-10M is a step toward that, fitted onto base models that were trained a different way. Those limits, and that longer aim, are Voltropy’s.

Early access is open. The blog’s closing line points readers to a signup for Vast-10M-Flash. The wire says people can sign up for early access to Vast-10M starting the day of the release at voltropy.com. The launch post links a technical report, and the wire says that report is available at voltropy.com. The pages do not print a price, a partner for the programming interface, or a customer logo.

In plain terms, an Orange County lab said on Tuesday that it shipped a three-size model family with a native window of ten million tokens, about ten times the one-million-token windows it says flagship models still advertise. The method is VSA. The BEAM numbers, including 40.20 at ten million tokens, are Voltropy’s. The signup the blog links is for the Flash size.

The picture is Voltropy’s Vast-10M launch graphic. A dark field carries the word VAST-10M and the line “The first frontier LLM with ten million tokens of context.” LLM is the short name for a large language model. A thin cream and gold frame sits around the card. It is the announcement image from the launch post. The card does not print a date.