Research snapshot · 20 September 2026
The mosquitoes win the tactic. The dinosaur wins the war.
Small, cheap, specialized models can beat much larger ones in the right arena. But the frontier labs have absorbed the same weapons—experts, parallel agents, search, and test-time compute—without giving up scale.
Two truths can coexist.
Compute at inference can let a smaller model punch far above its weight.
A compute-optimal small model can outperform a model fourteen times larger on problems where the smaller base model already has a non-trivial success rate.
Caveat: returns plateau at the base model’s capability ceiling. Test-time compute and pretraining compute are not interchangeable on the hardest problems.
Those same methods are now standard equipment for giant models.
DeepSeek-V3 holds 671B parameters but activates 37B per token: specialist experts inside one centrally trained model.
Caveat: “experts” here are subnetworks governed by one learned router—not independent, cheap agents.
The dinosaur didn’t fight the swarm. It internalized it.
External swarm
Many cheap models propose, critique, retry, vote, and route work to specialists.
Swarm inside the model
MoE routes each token to a few specialist subnetworks, keeping high capacity without activating the whole network.
All three values are from the DeepSeek-V3 technical report. The often-quoted “$5.6M training cost” is omitted here because it covers GPU-hours for that run—not hardware, salaries, failed experiments, or prior research.
Yes, xAI is a swarm operator. No, it isn’t evidence that mosquitoes replaced the dinosaur.
Half right—and that half cuts against the conclusion.
xAI uses swarm ideas at two levels: routed experts inside Grok-1 and parallel copies of a frontier model in Grok 4 Heavy. The second is a “study group” of dinosaurs, not a cloud of cheap mosquitoes.
Grok-1: sparse Mixture-of-Experts
314B total parameters; 8 routed experts; 2 active per token (about 86B, roughly 25% active). The 2024 open-weight release was a raw base model, not the production chat system.
Grok 4 Heavy: inference-time multi-agent
Multiple Grok 4 instances work independently, then compare and aggregate. Public reporting described Grok 4 as “single agent” and Heavy as “multi-agent.”
Unknown: Grok-2/3/4/4.5 internals
xAI has not publicly disclosed architecture or parameter counts for these models. Trillion-parameter and “four-persona” claims remain blog speculation.
The strongest case on both sides.
Open-model Mixture-of-Agents beat GPT-4 Omni on AlpacaEval 2.0.
65.1% versus 57.5%—a real ensemble win.
Wang et al. · verified primary ↗
But the “mosquitoes” were 70–110B-class models, and AlpacaEval is a GPT-4-judged preference benchmark—not a hard-reasoning test.
Self-MoA beat heterogeneous model mixing.
Sampling the single best model diversely delivered +6.6 percentage points over MoA on AlpacaEval 2.0 and +3.8 points on average across MMLU, CRUX, and MATH.
Li et al. · verified primary ↗
A strong monolith can simulate the productive part of a swarm.
Specialization changes the economics.
Narrow, well-specified tasks—classification, extraction, bounded QA—often reward fine-tuned small models. Private on-device deployment is also a structural advantage an API model cannot reproduce.
The frontier stays ahead on the hard tail.
Research summarized a roughly 3–4 month open-weight lag and 1.7× longer task horizons for the best closed models. The closed/open Elo gap reportedly widened from 0.5% to 3.3% in 2026.
Secondary frontier synthesis ↗
These are index/secondary figures, not independently reproduced here.
Serving cost is collapsing.
Equivalent-performance inference cost is reported to be falling by about 10× per year. Cheap inference is a real structural trend.
Secondary economics synthesis ↗
This does not uniquely favor small models: frontier providers benefit from the same hardware, batching, quantization, and MoE gains.
Naive debate can make weak agents worse.
Small-model debate has fallen below single-agent accuracy through sycophantic drift. The quality of the judge or verifier—not agent count—is often decisive.
Reframe the thesis, don’t throw it away.
Small systems will commoditize the middle and displace at the edges.
Most requests do not require frontier intelligence. Routing, specialization, local execution, and cheap retries can capture enormous volume.
Small systems will eliminate the frontier labs.
The likely equilibrium is coexistence: cheap systems handle volume; a few giant labs own the frontier, the hardest work, and the platform that rents swarm techniques back down.
The AT&T warning
SBC did acquire AT&T Corp. in 2005 and take its name. But the mechanism was antitrust breakup, deregulation, and M&A consolidation—not a cheap low-end swarm beating the giant.
The closer analogy
PCs displaced minicomputers because incumbents struggled to move downmarket. Frontier labs already offer mini/nano tiers, MoE efficiency, and open weights—so they can follow the market down.
Make economics—not raw intelligence—the swarm’s moat.
Cheap worker layer
Local and small models handle routine volume: button-pushing, retrieval, bounded transforms, monitoring, and repeatable tool calls.
Programmatic harness
Routing, retries, scoring, state checks, and distress signals supply the tenacity. The harness decides when a worker is outside its competence.
Intelligent boss
A strong coordinator handles ambiguity, synthesis, exceptions, and final judgment. Economize at the worker layer—not at the seat that evaluates the swarm.
This narrower thesis survives the evidence: the swarm need not be smarter than the frontier model. It must reserve expensive intelligence for the moments that require it, while cheap owned compute carries the continuous load.
The experiment that can kill or strengthen the idea.
Equal budget. Hard tasks. Binary scoring.
- Fix one hard, verifiable task family. Not preference judging—use outcomes with right/wrong answers and a hard tail.
- Fix the total inference-compute budget. Count every model call, retry, judge, and aggregation pass.
- Compare two systems. Best available swarm of genuinely small models versus one frontier model spending the same budget at inference.
- Publish the failure curve. Separate easy/medium cases from tasks where the small base models have near-zero initial success.
Snell et al. established the small-model win on problems where the small model already has non-trivial success. The unresolved question is the hard tail.
Evidence that would move the verdict toward mosquitoes
- A <10B open model matches the current frontier on HLE, ARC-AGI-2, or FrontierMath without distillation.
- The next 1–2 giant training generations fail to produce capability gains.
- A frontier lab exits giant-model training for economic reasons.
Evidence that would move it further toward dinosaurs
- Frontier capability jumps continue every 1–2 years.
- The open-versus-closed lag widens past roughly six months.
- Multi-agent gains continue to accrue most strongly to the largest base models.
What the public record still cannot settle.
Are xAI’s newer Grok models still MoE?
Unknown. xAI has not disclosed Grok-2/3/4/4.5 architecture or parameter counts. Grok-1’s MoE design does not prove continuity.
Do Grok 4 Heavy’s reported benchmark gains reproduce independently?
Not established in the research record. Public numbers such as 44.4% on Humanity’s Last Exam with tools were xAI-reported or repeated by secondary sources.
Will test-time compute keep scaling?
Unknown. More search and sampling help until the base model hits a capability ceiling. Whether reinforcement learning and inference-time search keep extending that ceiling is the central unresolved variable.
Could a genuinely small-model breakthrough change everything?
Yes. Current wins mostly come from specialization, distillation, orchestration, or spending more compute at inference. A <10B model matching the current frontier on hard reasoning without those crutches would materially change this verdict.
Four papers carry the core claims.
