Part of: this research — view all 6 rows
By late 2028, parameter-efficient adapters (principally LoRA) are expected to remain the dominant method for per-customer model customization by deployment count, driving a median estimate of roughly 18% for full-weight updates. The economics of multi-adapter serving heavily favor this outcome. Platforms like S-LoRA and Punica demonstrate that a single base model can host thousands of active adapters with roughly a 5% throughput penalty. Because of this, cloud and open-weight providers—including Google Vertex Supervised Fine Tuning for Gemini LLM, AWS Bedrock Reinforcement fine-tuning for Amazon Nova: Teaching AI ..., Together AI LoRA vs. full fine-tuning, and Fireworks Training API Now GA: Custom RL and Fine-Tuning Loops—all default to or exclusively offer adapter-based customization for their self-serve and serverless tiers.
Full-weight updates are largely confined to a small number of high-value, compute-intensive enterprise deployments, such as Mistral Forge Introducing Forge | Mistral AI and OpenAI's Custom-Trained Models Introducing improvements to the fine-tuning API and .... While Shopify’s August 2026 production loop successfully uses daily full-parameter fine-tuning Sidekick's continual learning loop (2026) How Shopify built a continual learning loop with ..., it trains a single shared model pooled across millions of merchants rather than maintaining per-customer customized weights, excluding it from this specific metric. Thus, full-weight commercial customization remains an exception reserved for organizations with substantial AI budgets and bespoke requirements.
The strongest technical argument for a shift toward full weights rests on continual learning. Research published in 2024 (arXiv 2410.21228) found that all LoRA ranks degraded faster than full fine-tuning across sequential tasks, suggesting adapters might struggle with continuous on-the-job learning. However, subsequent findings have tempered this view. A November 2026 Dialpad study https://arxiv.org/pdf/2609.27262 found that sequential LoRA actually preserved earlier skills better than sequential full-tuning in certain intent-classification tasks, indicating that adapters can remain viable even under continuous update regimes.
The ongoing deprecation of major self-serve fine-tuning platforms limits the path to widespread full-weight customization, concentrating that market into a smaller set of high-end enterprise deployments rather than encouraging broad adoption. This supports a median share of 18%, with the 10th and 25th percentiles at 5% and 10%, and the 75th percentile at 30%. The wide upper tail (p90 near 45%) primarily reflects market contraction and ambiguity in how the resolution might weight deployments. OpenAI's scheduled phase-out of all new self-serve fine-tuning jobs by January 2027 will remove a massive long tail of small adapter deployments from the market. While many of these users will likely transition to context-injection, RAG, or open-weight adapter platforms, the surviving customization market will be more concentrated among sophisticated buyers. That creates an upside risk through a smaller denominator, but does not itself imply growth in the number of full-weight deployments. Furthermore, if the resolution metric weights deployments by commercial spend or compute rather than a raw unweighted count, the handful of massive full-weight enterprise contracts would capture a dramatically larger share of the total.
Revised modestly downward to account for the ongoing deprecation of major self-serve fine-tuning platforms, which concentrates the full-weight market into a smaller set of high-end enterprise deployments.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited