Topic
Vendor and user shift toward compact models that trade parameter count for deployability, and the download patterns behind it.
Jeff, an open 0.8B model, returns a probability for each option in one forward pass and sends low-confidence agent decisions to Qwen3.8-27B. The confidence score attached to each answer tells an agent when a small decision is worth the larger model's time.
Reality
Qwen2.5-3B, wired to a local Wikipedia index, scored 52% on 150 post-cutoff questions it answers none of unaided, up from 33%, in a dev.to author's tests. Each fix targets a measured 3B failure, so a zero-shot 7B gained only 9 points from them, and the two readers' confidence intervals overlap.
Reality
Capsule Security has shipped an evaluator that runs inside an agent's execution path and can stop an intended action before it executes. Doing that requires adding a chokepoint to the agent's execution path.
Reality
The startup says Anchor 3.0 caught more than 90 percent of violations in a test it published itself, at under a 500th of the cost of a frontier call, and the messages it misses stay the firm's problem.
Reality
A case study reports an 8B model that got better at naming threat categories while losing ground on severity scoring, lifecycle counting and declining a weather question. Its author says every failure mode is already in the literature.
Reality
A Microsoft case study credits Docusign's swap to small task-specific models with 90% lower cost and eight times the throughput. Docusign's own accounting of the whole pipeline claims 50 times cheaper per document.
Reality
A KDnuggets walkthrough builds its comparison on Qwen2.5-0.5B-Instruct on an M2 MacBook Air, where the keys and values for a static instruction block come out identical on every call and only the ticket text changes.
Reality
ShadowPEFT is in Hugging Face PEFT's main branch as of a September 15th announcement, and it carries its own hidden state and can be detached as a smaller standalone model. Trying it means installing PEFT from source.
Reality
Microsoft's local runtime handles model download, hardware detection and execution provider selection behind a .dll, .so or .dylib. An app that adopts it owns the download, the cache and the unload.
Reality
The OpenAI bill it displaced was about EUR 10 a month. Paying that back takes a second set of calibration values, and the only comparison actually benchmarked in the writeup is embeddings.
Reality
A Llama 3.2 3B agent running offline invented a $1,990 balance on a $1,975 invoice. The open-sourced answer leaves prose to the model and hands every number to deterministic Python behind a tri-state router.
Reality
A 15M-parameter model streams English text on a 2007 PSP at about one token per second. That is the extreme end of a sizing rule. The harder half of that rule is checking whether the file that fits is a format its own maintainer recommends.
Reality
Stanford's AI Index puts a 142-fold parameter cut and a more than 280-fold price cut behind one fixed MMLU threshold. That narrows where building your own still pays, and it lands in a state-law count that doubled in a year.
Reality
Microsoft Research says a 4-billion-parameter model trained with SocialRL out-scored GPT-4.1, GPT-5.1 and GPT-5.2 across six bargaining domains. The winning margin is thinner than the method's effect.
Reality
A Fast Company column argues enterprise model selection is now cost optimisation. The consequence is that the layers deciding outcomes, context and feedback, are the buyer's own build.
Reality
Alibaba's Apache-2.0 Qwen3.8-27B fits in about 17GB and matched near-frontier scores, per Artificial Analysis. It also burned 3.7x the median output tokens getting there.
Reality