Voltropy Says Vast-10M Holds a Ten-Million-Token Context
Voltropy's 29 September 2026 post describes a 10 million token window and publishes its own BEAM scores against Fable 5.1 and GPT-6 Astra.
Direct answer
Voltropy's own post on 29 September 2026 says Vast-10M has a ten-million-token native context and that Vast-10M-Flash scores 40.20 on the ten-million-token BEAM tier.
On 29 September 2026 Voltropy published a product post for Vast-10M and described it as a model family with a ten-million-token native context. The post names the attention method VSA, Voltropy Scalable Attention, and says the family comes in Flash, Medium, and Pro sizes. Flash is described as based on DeepSeek V4.0 Flash, Medium on GLM-5.2, and Pro on DeepSeek V4.0 Pro. Early access, the post says, starts with Vast-10M-Flash.
The numbers below are the company's. At ten million tokens, the post says Vast-10M-Flash scores 40.20 on BEAM, against 39.69 for its base model at one million tokens. It says the Flash model keeps 82.49 percent of its one-million-token BEAM score at the longer tier. On the one-million-token tier it says Flash beats Claude Fable 5.1 by 4.70 points, is 10.68 percent higher than Fable, and reaches 97.95 percent of the score it attributes to GPT-6 Astra.
What the post does not establish
Those comparisons are not a third-party leaderboard. The post also says Vast-10M is not universally more capable than Fable or Astra on the hardest tasks, and it gives frontier mathematics as the example. The context claim is the product claim: a native window the post puts at ten times Fable 5.1 and 9.5 times GPT-6 Astra. Independent confirmation is not in the post.
Source: Voltropy, Introducing Vast-10M.
FAQ
- What does the post claim against Fable 5.1?
- It says Vast-10M-Flash beats Claude Fable 5.1 by 4.70 points on the one-million-token BEAM tier and is 10.68 percent higher than Fable at that length.
- Who measured those scores?
- The figures appear in Voltropy's product post. This article does not repeat them as an independent benchmark.