GLM-5.2 scores 99.2% on AIME 2026 with roughly 40 billion parameters active per token. Qwen3.5 scores 91.3% with 17 billion active.
For scale, GPT-4 was rumored to run around 280 billion active parameters in 2023, and it could barely solve an AIME problem at all.
Now ask the same models a plain factual question. On SimpleQA, a recall benchmark with no tools allowed, the leader is Gemini 2.5 Pro at 53%.
The best factual recall available still misses half the questions.
Why the Two Curves Move in Opposite Directions
The argument, laid out by developer Walter van der Giessen in a post that reached the front page of Hacker News this week, is that this is a deliberate trade rather than an accident.
Facts take space. Research on knowledge capacity, most cleanly measured in the Physics of Language Models series, puts it on the order of two bits of factual knowledge per parameter.
If you want a model that knows the birth year of every minor Wikipedia figure and the argument order of every function in every npm package, you pay for that in weights. That is a large part of why frontier models grew to trillions of parameters.
Reasoning compresses far better. It is a relatively small set of procedures applied repeatedly: break the problem into parts, track intermediate state, check the work, backtrack when a step fails.
So when a lab optimizes for capability per active parameter, reasoning survives the squeeze and factual recall does not.
| Model | Active parameters | Reasoning benchmark |
| GPT-4 (2023, rumored) | about 280 billion | Barely solved AIME problems |
| GLM-5.2 | about 40 billion | 99.2% on AIME 2026 |
| Qwen3.5 | about 17 billion | 91.3% on AIME 2026 |
| DeepSeek V4-Flash | about 13 billion | Not stated in the analysis |
| Gemini 2.5 Pro | Not stated | 53% on SimpleQA, the recall leader |
The Small-Model Result That Makes the Point
At the bottom end of the range, Qwen3.5 9B fits in 6GB of VRAM when quantized.
It roughly doubles the score of the next best sub-10B model on Artificial Analysis’s intelligence index.
A model that runs on a consumer graphics card now handles reasoning tasks that required a data center three years ago. It also knows dramatically less, and the two facts are connected.
What This Means Practically
The conclusion van der Giessen draws is that small local models paired with retrieval become more attractive than models with knowledge baked into weights.
If a model is going to miss half of factual questions regardless of size, paying for embedded knowledge buys you less than it used to. Better to run a small fast reasoner and give it access to a source of truth at query time.
That is an argument for retrieval-augmented generation over scale, and for local deployment over API calls, in a specific class of applications.
It is not universal. Tasks needing broad unprompted world knowledge, or long-tail context that is hard to retrieve, still favor larger models.
The Benchmark Caveat Worth Holding
SimpleQA measures recall of niche, tail knowledge with no tools. It also penalizes models that hedge appropriately, which is behavior you often want.
A model that says it is unsure rather than guessing scores worse on SimpleQA and is more useful in production.
So a 53% ceiling is a real limitation and a partly artificial one. The number should be read as a measure of unassisted recall, not of overall reliability.
Separate work on SimpleQA Verified has trimmed the dataset to reduce source-document overlap, which changes headroom and makes cross-version comparisons unsafe.
Why This Matters for Crypto and Fintech Builders
Anyone building an assistant over financial data has been implicitly betting on one side of this trade.
If your product answers questions about prices, filings, protocol parameters or regulatory status, embedded knowledge was never the right foundation. Those facts change, and a model trained six months ago is confidently wrong about them.
The reasoning-heavy, knowledge-light generation of models suits that architecture better. You supply the current data, the model does the inference.
Optimisus has covered the open-weight side of this shift, including Alibaba releasing a 2.4 trillion parameter model with open weights, where the active-parameter count was roughly 4% of the total.
The same reasoning gains underpin capability jumps in other domains, including the machine-checked mathematics proofs published earlier this month and the offensive security model released to vetted defenders.
The Honest Summary
These figures come from one analyst’s synthesis of public benchmark data, not from lab disclosures about training intent. No frontier lab has stated it is deliberately removing factual knowledge.
What is verifiable is the pattern. Reasoning scores climb while active parameter counts fall, and factual recall has not moved much in either direction.
The practical takeaway does not depend on intent. If you are choosing a model, benchmark it on the thing you actually need, because the headline number increasingly measures something narrow.
Sources
- Walter van der Giessen, Models are getting dumber on purpose — https://w4g1.dev/blog/models-are-getting-dumber-on-purpose
- AI Weekly, AI news today, August 17 — https://aiweekly.co/ai-news-today
- arXiv, SimpleQA Verified, a reliable factuality benchmark to measure parametric knowledge — https://arxiv.org/pdf/2509.07968
- BenchLM.ai, knowledge leaderboard and SimpleQA methodology notes — https://benchlm.ai/knowledge
Optimisus covers crypto and technology news for readers who want the detail behind the headline.
