Vast-10M is Voltropy’s model family with a claimed native context window of ten million tokens. The practical question is whether the extra input capacity improves the answers your application needs enough to justify its latency and cost. A context limit alone cannot answer that.
Voltropy’s September 29 announcement describes three variants built by modifying existing models with Voltropy Scalable Attention, or VSA. Flash uses DeepSeek V4.0 Flash, Pro uses DeepSeek V4.0 Pro, and Medium uses GLM-5.2. These are developer claims about the released family, rather than independent EyesTech measurements.
What the reported scores show
The technical report lists Flash at 48.73 on the 1M-token BEAM tier and 40.20 at 10M. It says the full tiers were evaluated: 700 questions at 1M and 200 at 10M, with the official GPT-4.1-mini judge at temperature zero.
The report also cautions that tiers contain different conversations and questions. Their scores are useful comparisons, but do not describe identical tasks repeated with longer filler. Tokenization sometimes required minimum-prefix truncation for models to accept a nominal tier. Those details limit how literally readers should interpret a retention percentage.
The score ratio is 40.20 / 48.73 × 100, approximately 82.49%. The absolute decline is 8.53 score points. These calculated descriptions of the same endpoints answer different questions. Neither establishes that an application preserves 82.49% of every fact when its prompt grows tenfold.
The original BEAM research concerns memory abilities over long, coherent conversations. It introduces 100 conversations and 2,000 validated questions. That is a more specific target than arbitrary source-code repositories, regulatory archives or image-heavy documents. An application evaluation should preserve its own document structure and question types.
Test the value of extra context
A useful comparison starts with a frozen set of questions and source documents. For each question, record the required evidence and an answer rubric before calling either model. Keep the same evidence available when comparing full context with retrieval, and retain the exact text each system actually receives.
For example, construct a synthetic project archive with a January decision, a February revision and a March exception. Ask for the current decision and the reason it changed. A correct answer must connect the revisions, cite the right records and avoid applying the exception outside its stated conditions. Repeat the question after adding unrelated material in known positions.
This proposed procedure keeps the answerable task stable while changing the amount and placement of distracting text. It also allows failures to be classified: evidence omitted from the request, evidence present but overlooked, an obsolete rule selected, or an unsupported answer generated. A single final score hides those differences.
Measure time to the first useful answer, total wall-clock time, actual billed tokens and citation accuracy. If a long request fails and must be retried, include that retry in the cost of the completed task. If the provider caches repeated input, report cold and warm runs separately rather than assuming every request receives the same benefit.
Native context does not settle the retrieval decision
A large window offers a reason to evaluate sending more source material directly. Retrieval remains a candidate when a system needs access controls, frequent document updates or a traceable evidence selection step. The better choice depends on the failure modes and operating requirements of the particular workload.
An illustrative cost model makes the tradeoff explicit. If an uncached request contains ten million input tokens and the input price is P dollars per million, its input charge alone is 10P dollars. That excludes output, retries and any separate tool charges. This is arithmetic with a symbolic price, not a quote for Vast-10M.
A provider evaluation therefore needs both the endpoint’s current limits and its commercial terms. A benchmark PDF cannot establish the price, rate limit or service commitment of a future deployment.
How to get access
Voltropy’s current early-access page presents a waitlist form and says it will email when access is ready. That page does not establish immediate self-service access for every applicant. We did not submit an application or verify a live account.
Before integrating, request the exact model identifier, supported input and output limits, billing rules, data retention terms and a way to reproduce the benchmark configuration. The evidence that would materially strengthen the deployment case is an inspectable run on your fixed question set, with the complete input, scored answers and operating costs retained.
Vast-10M’s reported long-context results justify that evaluation. They do not yet determine whether replacing retrieval or another model improves a particular production task.
Reporting note: Prepared with AI assistance from the linked primary sources. Arithmetic and proposed evaluation procedures are labeled; EyesTech did not run an independent Vast-10M benchmark. The featured image is an original AI-generated editorial illustration.
