# Unwrapping the wrappers > Canonical: https://www.overmindlab.ai/research/who-trains-their-own-models - Author: Sam Brunt - Category: Research - Tags: Models - Published: 2026-07-30 - Updated: 2026-08-07 > Application layer companies are training their own models. Why? ![Unwrapping the wrappers](https://cdn.sanity.io/images/gv4t9qa3/production/52436243946c705c3e53d056b89da88ec7764636-2048x2048.png) Application layer companies, often labelled as 'ChatGPT wrappers', are starting to train their own models. Specialist models, post-trained or fine-tuned on a dataset their competitors cannot touch: their own production traffic. Four recent examples: - **Cursor** trained Composer, the model inside its coding agent, with online RL on production agent traces. Delivering frontier-level coding at a fraction of the latency and cost. - **Intercom** trained Fin Apex, a customer-service model post-trained on billions of support interactions. It bought them control of the resolution-rate metric at scale. - **Replit** trained a 7B code-repair model, a fine-tune of DeepSeek-Coder-Instruct-v1.5. It bought them a fast, narrow specialist trained on IDE data nobody else has. - **Perplexity** trained Sonar, its search-answer model, a fine-tune of Llama 3.3 70B served on Cerebras hardware. It bought them sub-second, citation-heavy answers at search scale. ## Cursor: reinforcement learning on live agent traces Cursor didn't fine-tune once and ship. The Composer family of models run on real-time RL: model checkpoints go to production, real user interactions become the reward signal, and an improved checkpoint ships several times a day. The model is a mixture-of-experts architecture trained inside a harness built to mirror Cursor's actual tool space, including file edits, semantic search, and terminal commands. The reward signal reflects how developers really use the product rather than a synthetic benchmark, according to [Cursor's technical post on Composer](https://cursor.com/blog/composer). Cursor reports Composer generates tokens several times faster than comparable frontier models while landing close to the top tier on internal coding benchmarks, a tradeoff explored in more depth by the [Composer 2 technical report](https://arxiv.org/html/2603.24477v1). Nobody else has Cursor's trace data, which means nobody else can train Cursor's model, even with an identical base checkpoint and an identical training recipe. What it bought them: coding performance competitive with larger models, at preferential latency and cost, plus a moat provided by their own proprietary data. ## Intercom: post-training on the metric the product is sold on Fin launched in 2023 on top of GPT-4. Fin 2 moved to Claude in late 2024. Then Intercom's roughly 60-person AI research group, led by Fergal Reid, built Fin Apex 1.0, a model post-trained on the company's own support interaction data, as described in the [Fin Apex launch post](https://www.intercom.com/blog/announcing-fin-apex-the-age-of-vertical-models-is-here/). One large enterprise customer saw its resolution rate jump from 68% to 75% overnight after the switch to Apex, a 22% cut in unresolved conversations according to Intercom's own numbers. Independent coverage from [VentureBeat](https://venturebeat.com/technology/intercoms-new-post-trained-fin-apex-1-0-beats-gpt-5-4-and-claude-sonnet-4-6) puts Apex's benchmark resolution rate at 73.1%, against 71.1% for both GPT-5.4 and Claude Opus 4.5. Intercom's own model documentation breaks the pipeline into separate retrieval, reranking, and generation models rather than one monolithic system, detailed on the [Fin CX models page](https://fin.ai/cx-models). Worth noting: Intercom hasn't abandoned rented models. Apex is layered on top of retrieval and routing components, and the company still uses third-party models elsewhere. This architecture is common as companies balance performance, latency and cost. ## Replit: a narrow 7B specialist trained on IDE telemetry Replit's bet was the opposite of "bigger is better." The team fine-tuned a 7B open-weights model, DeepSeek-Coder-Instruct-v1.5, chosen after testing several code-focused base models and documented in its [code-repair write-up](https://replit.com/blog/code-repair). The model targets a specific use case, roughly the 20 diagnostic categories responsible for the majority of everyday coding errors, and the fine-tuned 7B checkpoint matched or beat GPT-4 and Claude 3 on real-world fixes in that use case. Small [open source models](https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5) can beat much larger general-purpose models at many tasks. ## Perplexity: search-optimised inference at 1,200 tokens per second Perplexity fine-tuned Meta's Llama 3.3 70B into Sonar, tuned specifically for factual grounding, citation quality, and readability in its default search mode, as the company explains in its [Sonar launch post](https://www.perplexity.ai/hub/blog/meet-new-sonar). Sonar runs on Cerebras wafer-scale inference hardware, which Perplexity and [Cerebras jointly announced](https://www.cerebras.ai/press-release/cerebras-powers-perplexity-sonar-with-industrys-fastest-ai-inference) pushes throughput to roughly 1,200 tokens per second, an order of magnitude faster than comparable models on conventional GPU inference. In Perplexity's own A/B testing, Sonar matched or beat larger frontier models on user satisfaction inside search, while running far cheaper and faster than calling those models directly. ## Why build a specialist model instead of prompting a frontier one? Calling a frontier model for every request is sometimes the equivalent of running a data centre to toast bread. It works, but it's the wrong tool for a narrow, high-volume, well-defined job. Across these four companies, three conditions kept showing up: - **A narrow task, not a general one.** Small specialised models are excellent at one narrow task. Don't ask them to be broadly intelligent. - **Proprietary production data no competitor can access.** Cursor has Cursor's traces. Intercom has Intercom's support conversations. Your production traffic is the one dataset nobody can buy. - **Enough volume that per-token latency and cost actually move the P&L.** Fine-tuning a model can be expensive. It only pays off once inference volume is high enough that shaving cost per call compounds. None of these companies want to out-build OpenAI or Anthropic on general intelligence. Each trained a model that wins one specific task, using the one dataset it uniquely owned. ![Flywheel diagram titled "Traces into training". Production runs feed a labelled dataset, the labelled dataset feeds a fine-tuned model, and a dotted redeploy loop sends the fine-tuned model back to production runs.](https://cdn.sanity.io/images/gv4t9qa3/production/e39a777c3f8b8c4038ebc85b5a5db9ad5f78b6c4-2014x1012.png) Capture production traces, turn them into a labelled dataset, train a model on it, and ship that model back into production. That's where the application layer is going. _Overmind is the model training platform for AI teams. It turns your production traces into specialised models you own. [Get started](https://console.overmindlab.ai)._ ## Related research - [When bigger isn’t better](https://www.overmindlab.ai/research/when-bigger-isnt-better): On specialized tasks, Overmind-trained small language models outperform frontier models on accuracy, hallucination, and cost. - [How to train your agent](https://www.overmindlab.ai/research/how-to-train-your-agent): Learnings from building, deploying and improving agents in production. - [How do you turn traces into a training dataset?](https://www.overmindlab.ai/research/traces-to-training-dataset): Every agentic system can create its own training data.