Artificial-intelligence models are becoming better at structured reasoning while retaining less detailed factual knowledge in their weights, according to an analysis published on August 16, 2026. Developer Walter van der Giessen argues that this is a deliberate and potentially useful trade: procedures can be compressed into smaller models, while facts can be supplied when needed through search, documents and other tools.
The essay contrasts mathematical performance with unaided recall. It reports GLM-5.2 scoring 99.2 percent on AIME 2026 while activating about 40 billion parameters per token, Qwen3.5 scoring 91.3 percent with 17 billion active parameters and DeepSeek V4-Flash operating with 13 billion active parameters. Those figures are presented by the author from benchmark sources referenced in the essay; they were not independently tested in the supplied evidence.
Factual benchmarks show a different pattern. The essay says Gemini 2.5 Pro led SimpleQA at 53 percent, while Artificial Analysis measured hallucination rates of 80 to 82 percent for Qwen3.5 models with four billion and nine billion parameters on its knowledge test. The author’s interpretation is that small models can learn reusable methods for mathematics and coding without storing the long tail of names, dates, software versions and other facts.
That distinction matters because stored knowledge becomes outdated. Prices, employment, libraries and product interfaces can change before a lengthy training cycle finishes. Algebraic methods and techniques such as decomposition, checking work and backtracking are more stable. Moving current facts outside model weights could therefore separate the expensive trained system from information that changes daily.
In this model of AI use, the surrounding harness becomes essential. A coding agent can inspect installed dependencies, search documentation or read files rather than relying on an API version remembered from training. Retrieval changes factual recall from a cost paid on every model invocation into an on-demand operation grounded in the user’s actual environment. It can also make an error traceable to a document that can be inspected and corrected.
External knowledge does not eliminate hallucinations. A model may misunderstand a source or combine material incorrectly, and a tool-free request may still produce a confident mistake. The essay argues that a well-designed system should instead acknowledge uncertainty and retrieve evidence. That behaviour depends on the harness, instructions and accessible information, not only the model.
The author projects that models with strong reasoning and total sizes closer to their active parameter counts could eventually run on a consumer GPU, particularly with four-bit quantization. That remains a forecast rather than a demonstrated outcome. The present evidence supports a narrower conclusion: benchmark efficiency is improving, but reasoning scores alone do not show that a model has broad, reliable factual recall.
