Skip to main contentSkip to footer
Small Models Answer Questions. Big Models Get the Job Done.
HomeNewsExpert NoteLarge AI Models Are Back

Large AI Models Are Back

Small Models Answer Questions. Big Models Get the Job Done.

After two years of “small is smart,” AI agents are changing the equation. When AI has to complete long, multi-step tasks on its own, model size makes the difference, and Swiss organisations can now run that power on sovereign infrastructure.

Act 1 (2023): bigger LLMs are better

In 2023 the playbook was trivial: call the largest model behind an API and build around it. Nothing else came close. By the end of that year OpenAI alone held 50% of the enterprise LLM market (Menlo Ventures).

The use cases were also simple: a chatbot, a summary, a first draft. One prompt in, one answer out, a human checking the result.

Act 2 (2024 to 2025): the rise of small language models

Then the bills arrived, and with them a new narrative. Distillation, pruning and better training data let compact models punch far above their weight. Microsoft's Phi-4, at 14B parameters, beat GPT-4o-mini on math and code benchmarks. (Meta Intelligence).

 

The arguments were real:

 

  • Cost and latency: one tenth to one fiftieth of the deployment cost of a large model for many tasks.
  • Privacy and control: run it locally, keep data in-house.
  • Good enough: most enterprise NLP is classification, extraction and summarization.

 

The narrative peaked with NVIDIA's 2025 position paper arguing that small models are the future of agentic AI, and Gartner forecasting a threefold rise in task-specific models by 2027 (HackerNoon, Medium). Analysts still label 2026 the year of efficiency (Tech Monitor)

Act 3 (2026): the pendulum swings back to large models

The value moved from answers to autonomous work, and autonomous work favors scale. METR measures how long a task (in human expert time) an AI agent can complete on its own. That horizon has been doubling roughly every four months since 2023 (METR).

 

  • Claude Opus 4.6: about 12 hours of expert work at 50% reliability (AI 2027 Tracker).
  • Claude Mythos: at least 16 hours at 50%, and just over 3 hours at 80% reliability (METR).

 

Every model at the top of that chart is a frontier-scale model.

The clearest signal comes from the open-weight world, the camp that once championed small. In six months it shipped the largest models ever released:

 

ModelReleasedTotal / active parameters
Kimi K3 (Moonshot AI)July 20262.8T / ~100B
GLM-5.2 (Z.ai)June 2026~744B / 40B
DeepSeek V4 ProApril 20261.6T / 49B

 

GLM-5.2 landed fourth on the Artificial Analysis Intelligence Index, ahead of every closed model you cannot download (Tech Insider). The gap between closed and open frontier has shrunk to three to five months (AI Weekly). Nobody closes that gap with a 7B model.

What open-weight models achieved in 2026

You do not need a closed model to see the gap. All the examples below come from open-weight models.

 

  • Cyber offense, next to the gated frontier. GLM-5.3 (Z.ai) built end-to-end exploits in 50 of 410 ExploitBench attempts, against 56 for Anthropic's gated Claude Mythos Preview; earlier models scored near zero (AI Weekly). Its smaller Flash variant turned a public Chrome flaw into a working exploit chain with 8 hours of model work and about $20 of inference. NIST's CAISI rates GLM-5.3 the most cyber-capable open-weight model to date (NIST).
  • A chip designed in 48 hours. In a single 48-hour autonomous run, Kimi K3 built, optimized and verified a chip using open-source EDA tools (Kimi K3 report). Moonshot also showcases a Triton-like compiler built from scratch, and an astrophysics reproduction cut from one to two weeks of expert work to about two hours (Alexandre Agius). These are vendor showcases.
  • Olympiad gold for 12 cents. In an independent run on IMO 2026 problems, DeepSeek V4 Flash scored 30/42, above the 29-point gold cutoff and nearly twice the human median, for $0.12 (Cline). Proofs were graded by model judges, not the official jury.
  • Real-world coding. GLM-5.2 scores 62.1% on SWE-bench Pro, ahead of GPT-5.5 at 58.6% (DeepInfra).

 

Look at the sizes: GLM-5.2 is about 744B parameters, Kimi K3 is 2.8T, and even DeepSeek's "Flash" is 284B. Open source did not catch the frontier by going small; it caught it by going big. GLM-5.3 also cuts both ways: Anthropic stripped its refusals for a few thousand dollars of compute, so whoever hosts it carries the governance (Pasquale Pillitteri).
 

Why small models cannot do the same

An agent's success is the product of every step it takes. A model that is right 99% of the time per step completes a 100-step task only about a third of the time. Tiny per-step differences, invisible on a single-question benchmark, become a gulf on real work.

 

Per-step accuracy10 steps50 steps100 steps
98%82%36%13%
99%90%61%37%
99.9%99%95%90%

 

Research confirms the effect is real, and worse than simple arithmetic suggests:

 

  • Larger models execute far more steps correctly, even when small models show near-perfect single-turn accuracy (Sinha et al.).
  • Models degrade by conditioning on their own earlier mistakes; once an error enters the context, more follow (Sinha et al.).
  • Across 23,392 agent episodes, success dropped from 76% on short tasks to 52% on very long ones, and model rankings reversed between the two (Tian Pan).


That last point is the trap. A small model that ties a big one on a leaderboard can still lose badly on a two-hour task. Benchmarks measure the flat start of the decay curve; production lives at the end of it.
 

The cost illusion

The right unit is cost per completed task, not cost per token. A model ten times cheaper per token that finishes 13% of long tasks instead of 61% is not a saving. Add the retries, the human who supervises, and the silent errors nobody catches until production.
 

The price argument is also eroding from the other side. Near-frontier open models such as DeepSeek V4 already run 10 to 13 times cheaper than the top closed APIs (Tech Insider). Buyers have noticed: by mid-2025, closed frontier models powered 87% of enterprise workloads and open-source adoption had flattened (BigDATAwire).
 

Where small models still belong

Small models are not dead; they are being demoted from brain to specialist. They remain the right tool for narrow, high-volume, well-defined tasks:

 

  • Classification, routing and entity extraction
  • On-device and offline use, from phones to industrial equipment
  • Sub-tasks inside an agent, called by a larger orchestrator


That last pattern is becoming the enterprise standard. As AT&T's chief data officer puts it, large reasoning models run the control loop while fine-tuned small models execute dedicated steps (Mobile World Live). The architecture is tiered, and the top tier is big.
 

What this means for European and Swiss enterprises

The sovereignty question is no longer small and local versus big and American. Much of the small-model enthusiasm in Europe was really a sovereignty argument: a model you can run yourself beats a model you must send your data to. That trade-off made sense in 2024. It no longer does.

 

Frontier-class open weights now ship under MIT-style licenses. The constraint has moved from the model to the infrastructure: these models need multi-GPU nodes or clusters, not a server in the basement (Yotta Labs).

 

For a bank, a hospital or a canton, sovereign AI in 2026 means one thing: frontier-scale open models running on GPU capacity inside the country, under local law. Settling for a small model to stay sovereign means accepting an agent that fails most of the real jobs you give it.
Small models will keep winning tasks. Big models are winning the jobs.
 

Ready to move from AI answers to AI that gets the job done?

Choosing the right model is now a strategic decision, not just a technical one. EveryWare, the cloud specialist within ELCA Group, will help you identify where AI agents can create real value in your organisation, design the right mix of models, and run them securely on Swiss infrastructure.