Open model usage is concentrated in a few practical areas, with creative writing, role-play, and companionship making up the largest non-coding category.
2
Production users keep older models because stable behavior, existing prompts, tutorials, and measured reliability matter more than switching to the newest benchmark winner.
3
Enterprise agents should start by automating about 80% to 90% of a workflow with human escape hatches, then expand gradually instead of aiming for full automation immediately.
Summary
Eugene Cheah examines how people use open-weight models through usage data from Featherless.ai and comparisons with OpenRouter. Individual users favor DeepSeek R1 and other well-known models, while commercial workloads remain attached to older models such as Mistral Nemo and Llama 2 because teams value stable versions, existing integrations, and predictable behavior. Creative writing, role-play, companionship, and related uses account for the largest share of non-coding traffic. Coding represents roughly 20% to 30% of traffic, with auto-completion largely solved and agentic coding becoming the main area of competition. Cheah also describes image and creative workflows, ChatGPT-style interfaces, and enterprise automation. His advice for agents is to launch with human review, automate reliable portions first, and expand through repeated iterations. The closing segment presents Featherless's Quirky model and argues that benchmark scores matter less than whether an agent can complete one task reliably.
Open model supply is growing quickly, but downloads do not explain usage by themselves
Cheah opens with the scale of the open model ecosystem. More than 50,000 models had reportedly been uploaded to Hugging Face per month, which he frames as more than one model every minute. DeepSeek R1 had more than 4 million downloads in the prior month, and its model files represented a huge amount of network traffic. He uses this growth to ask what people actually do with open models. Featherless.ai provides access to more than 3,700 open models through a flat individual plan, giving Cheah usage data based on selected models and requesting applications.
Individual users choose models by familiarity and feel when price is held constant
For a week in February 2025, individual users on Featherless.ai favored DeepSeek R1, followed by Llama-family models, Mistral Nemo, and Qwen models. These users were not charged by token. The individual plans allowed one large model request at a time, with smaller models usually responding faster, and did not impose a price difference between models. Cheah says this makes people choose according to preference, reputation, and what he calls "vibes" rather than benchmark scores or cost. Looking at model names instead of broad model classes also reveals how fine-tuned versions of Llama and Qwen split into models with distinct personalities and use cases.
Older models remain in production because teams value consistency over upgrades
Commercial workloads change the picture. Mistral Nemo, an older and smaller model, remains heavily used even though newer and larger models are available. Cheah attributes this to lower cost at scale, production stickiness, and its early adoption under a license that enterprises found more comfortable than the Llama license. Cloud providers promoted Mistral Nemo as a replacement for GPT-3.5 workloads, and many tutorials and fine-tunes were built around it. Once a business has prompts, monitoring, and a reliable system in place, changing the model can break behavior. The same pattern keeps Llama 2 and Llama Guard in use, including for new teams following existing safety tutorials.
Creative writing, role-play, and companionship are the largest non-coding use case
Cheah groups creative writing, role-play, companionship, and some therapy and journaling under a broad category that accounts for roughly 30% to 40% of AI traffic outside coding. Examples include NovelCrafter for planning and drafting novels, and apps such as Wyvern and Character.AI-style services for role-play or companionship. He says this is the largest non-coding, non-agent use case by active users, despite the stigma around some of these applications. More than 60% of users in this segment are women, according to the data he presents. These users care less about benchmark scores and more about conversational style, pacing, and the particular personality of a fine-tuned model.
Vibes-based applications reward slow, distinctive conversations rather than fast answers
In therapy-style conversations, the model is expected to guide a person toward their own decision instead of simply telling them what to do. In fiction, role-play, and companionship, users want a "slow burn" rather than a shortcut to the ending. This creates demand for many community fine-tunes with different styles. Cheah names Magnum as a cowboy-oriented model and Rosé as a model focused on science fiction, while acknowledging that popular models change quickly. Users may still return to an older favorite because it feels familiar. He estimates that this wider community, including closed commercial apps such as Character.AI, has tens of millions of users.
Coding traffic is shifting from autocomplete to agents with people in the loop
Coding copilots and coding agents account for roughly 20% to 30% of traffic. Cheah considers basic code completion largely solved because many small models, from 3 billion to 12 billion parameters, are good enough for the task. The competition has moved to agentic coding, where developers chat with a system, answer clarifying questions, and intervene when it goes wrong. He calls this style "vibe coding." These workflows can generate far more token traffic than ordinary chat. Claude Sonnet still dominates the cited OpenRouter data by more than 10 to 1, but open model usage has grown since the DeepSeek R1 release. Cheah recommends trying open projects such as Cline and Continue with tools such as VS Code.
Visual workflow tools are spreading beyond developers
ComfyUI and similar graph-based interfaces account for about 5% of traffic in Cheah's data. They began in image generation, where users chain several operations into a larger workflow, but the same style is being adopted by reporters, lawyers, musicians, influencers, and other creative workers. These users are not necessarily developers. They learn the tools to build workflows for their own work. Cheah describes this market as a framework competition that has grown gradually over recent months, rather than showing the sudden rise he associates with some coding tools after DeepSeek R1.
Enterprise agents should automate a reliable portion of work before removing human review
Cheah divides agent use into workflow automation with human controls and fully automated agents. For an insurance or logistics email process, an agent can draft replies, check inventory and rules, and present the result in an interface where a person edits, approves, and sends it. He reports that a system can draft about 80% to 90% of responses at launch, with humans handling the remainder. Teams can later automate specific scenarios after observing the system across many customers. Projects that begin with full automation can fail after one damaging response, which may delay wider AI adoption. Cheah recommends repeated incremental improvement across failure cases, with human escape hatches built in from the start.
"Most developers want consistency. They want to only change their model when they choose to do so, or opt in to do so, not when their provider decided to do an update."02:56
Who should watch
You are choosing open models for a production system and need to understand why an older model may be safer to keep than a newer one.
You are building a coding agent and want a view of where autocomplete ends and human-in-the-loop agent workflows begin.
You are responsible for an enterprise automation project and need a rollout approach that does not depend on perfect autonomous behavior.
