Directions Brain

slm-default-and-fine-tuning

3 min read

SLM Default and Fine-Tuning

Definition

The thesis that small, domain-tuned language models will be the default enterprise inference path by 2027, displacing general-purpose frontier LLMs for the workflows where domain specificity matters more than raw capability. Cited in 2026-05-01-entree-capital-enterprise-ai-spend-map from Gartner (via InfoWorld, May 2026): SLM use 3x larger than LLMs in enterprise by 2027. Pairs with vertical-use-case-led-brain as the infrastructure-side moat: pick a vertical, build a domain corpus, fine-tune your own SLM on it, and the resulting model becomes a defensible asset that the frontier API can’t replicate without your data.

Key points

  • The Gartner 3x prediction. SLM enterprise use ≥3x LLM enterprise use by 2027. Direct implication for any vertical play: domain-tuned smaller models are the medium-term default, not the exception.
  • Pie-share evidence inside the report. Self-hosted models 6% (2025) → 10% (2030); fine-tuning 2% → 4%. Both shares roughly double while the total pie 5x’s, so absolute self-hosted + fine-tuning grows ~10x.
  • Why SLMs win in vertical contexts. Three reasons the report and analysts cite: (a) cost — token economics on a small fine-tuned model are an order of magnitude cheaper than frontier API calls; (b) latency — small models run inside the request-response loop; (c) domain correctness — a model fine-tuned on a vertical’s actual corpus outperforms frontier-general models on that vertical’s tasks.
  • The moat argument. If the team picks a vertical (e.g. healthcare clinics in clinical-data-portability), the durable asset is not the wrapper UI — it’s the domain corpus and the SLM trained on it. Frontier labs can’t replicate this without access to the same domain data. This is the layer where vertical-use-case-led-brain becomes defensible against openai / anthropic eventually shipping the same vertical surface.
  • Counter-argument the team must take seriously. Frontier models keep getting better at zero-shot vertical tasks. Every six months, the floor of “good enough without fine-tuning” rises. An SLM moat that holds in 2026 may not hold in 2028. The hedge: build the domain corpus first (which is durable) and treat the SLM as the current best way to monetize it — not as the moat itself.
  • Practical tooling implication. Fine-tuning shifts from a 17B category by 2030 per the report’s stacked-bar chart. Fine-tuning APIs (openai / anthropic FT, AWS GPU) are the on-ramp; self-hosted on AWS / Azure / CoreWeave is the destination.

Evidence

Open questions

  • Is the right move for a vertical wedge to ship on frontier API initially (faster TTV) and migrate to a fine-tuned SLM only after PMF? Or is the moat lost if the corpus accumulates inside the frontier provider’s logs?
  • Which vertical the team picks has the most defensible domain corpus? Healthcare clinical data is famously hard to legally aggregate; legal (Harvey’s playbook) is easier; call-centers (notch / wonderful pattern) is in between.
  • Does Monday’s monday-agents-month generate a corpus the team could use, or is each agent’s data too narrow / Monday-specific to fine-tune anything broader?
  • Open-weight vs closed-weight SLMs: Llama / Mistral / DeepSeek / Qwen on AWS/Azure/CoreWeave is the reported default — but open-weight licenses for fine-tuned commercial use are getting messier. Worth tracking.

Graph View

Backlinks