Vice President · TheGreyMatter.ai
In brief
Recent releases expand the shortlist. Workflow evidence decides what belongs in production.
A model announcement can change what a portfolio company should test. It cannot establish what the company should deploy. The useful question is which part of a real workflow a release might improve, and what evidence would demonstrate that improvement after human review.
Two releases, two different jobs
IBM announced Granite 4.2 on August 25, 2026, in 3B, 8B, and 30B parameter sizes. IBM describes these generative models as supporting reasoning and agent workflows. Those are vendor descriptions, not results from a portfolio company evaluation. The range also illustrates why the label small needs context rather than a universal cutoff.
Google lists EmbeddingGemma 2 as an October 6, 2026 release. Its model card describes a 740M parameter model that represents text, images, video, and audio in a shared vector space. That makes it relevant to retrieval. It does not replace the generative model that writes an answer or prepares a decision brief.
Start with a document question
Consider a proposed operating review assistant. Its first job might be finding the current definition of recurring revenue and the approved monthly figures. Its second job might be explaining a variance with citations. Evaluate those jobs separately. A fluent explanation cannot compensate for retrieving last quarter’s definition or a document from the wrong company.
Create a small evaluation set from authorized material. Include conflicting versions, incomplete evidence, uncommon terminology, and questions the system should decline to answer. Have an operating owner identify the relevant passages before comparing systems. Keep a separate set of cases for the final decision so prompt tuning does not become rehearsal for the test.
Compare the whole operating bill
Record infrastructure, indexing, monitoring, support, and reviewer time alongside inference charges. Measure the time to an accepted output, including corrections and escalations. A smaller model that generates more rework can cost more per useful result. A larger model may justify its cost on difficult cases while a focused model handles routine extraction.
Private deployment needs an actual boundary diagram. Identify where documents, embeddings, logs, backups, and support access reside. Check which services receive requests and who controls each identity. Parameter count alone tells you nothing about those paths, and running locally does not automatically resolve retention, access, or recovery requirements.
Roll out one comparable workflow
Across a portfolio, reuse the evaluation method before reusing the answers. Companies may define churn, customers, and pipeline differently. Preserve those definitions and permissions when comparing results. A shared scorecard should distinguish retrieval quality, accepted output quality, access failures, latency, and total cost rather than compressing them into one impressive number.
The next step is a bounded pilot with an owner, a baseline, and a stop condition. Record the model version and configuration so later releases can be compared honestly. Expand only when the workflow remains useful under normal exceptions. The value of a new release is the additional option it creates, not a promise that the operating work has disappeared.
Sources and evidence
AI assisted editorial analysis. Sources checked October 10, 2026. Practical examples are proposed workflows, not reported customer outcomes or personal experience.
One useful next step: Choose one idea from this note and test it at the smallest scale that could teach you something this week.
