PewDiePie launches Ajax AI model based on Qwen3.5-9B for private agents running on home PCs
PewDiePie has released Ajax, a 9-billion-parameter AI model built from Alibaba's Qwen3.5-9B and specialized to operate as a local agent inside Odysseus, his open-source, self-hosted AI workspace.
Rather than competing with frontier models primarily on general benchmarks, Ajax targets a narrower deployment model: an always-available personal agent running on hardware controlled by the user, with access to tools for web research, files, email, calendars and other everyday workflows.
Ajax has also been deliberately modified to reduce the refusal behavior inherited from its base model. PewDiePie describes the resulting system as "uncensored," although the project retains a stated boundary around requests involving harm to the user or other people.
The release combines three increasingly important trends in consumer AI: small open-weight models, local inference and agents capable of acting on private personal data without requiring every interaction to pass through a hosted frontier-model provider.
··········
AJAX AT A GLANCE
........
Specification | Ajax |
Developer | PewDiePie / Felix Kjellberg |
Base model | Qwen3.5-9B |
Parameters | 9 billion |
Primary role | Local personal AI agent |
Main environment | Odysseus |
Deployment | Self-hosted / local |
Agent capabilities | Tools, web, files, shell, skills and memory |
Personal workflows | Email, calendar, documents, notes and tasks |
Refusal modification | Abliteration using Heretic |
Model philosophy | Small, specialized and locally deployable |
Workspace model support | Local models and external APIs |
Odysseus license | AGPL-3.0-or-later |
........
Ajax should be distinguished from Odysseus itself.
Ajax is the language model while Odysseus is the surrounding agent workspace that provides interfaces, tools, memory and connections to other applications.
That separation allows Odysseus to work with Ajax locally while retaining support for other local models and external model APIs.
··········
AJAX STARTS FROM QWEN3.5-9B RATHER THAN A FRONTIER-SCALE MODEL
Ajax is based on Qwen3.5-9B, placing it in a very different compute class from frontier systems containing hundreds of billions of total parameters.
The choice is intentional.
A smaller model requires less memory and compute, making local inference practical on substantially more hardware. It can also reduce latency created by repeated communication with remote APIs and eliminate per-token API charges when inference is performed entirely on the user's machine.
The trade-off is model capability.
A 9B model cannot be assumed to match substantially larger frontier systems across difficult reasoning, coding, knowledge and long-horizon agent tasks. Ajax instead attempts to compensate through specialization around a particular agent environment and its tools.
For a personal agent, successful tool selection, structured actions and reliable interaction with the surrounding workspace can matter as much as maximizing performance on broad language-model benchmarks.
··········
ODYSSEUS TURNS THE MODEL INTO A PERSONAL AI WORKSPACE
Ajax is designed around Odysseus, PewDiePie's existing self-hosted AI environment.
Odysseus provides much of the functionality that turns a language model into an agent rather than a standalone chatbot.
The workspace supports chat, autonomous agent execution, files, shell access, MCP integrations, skills and persistent memory. It also includes interfaces for deep research, documents, email, notes, tasks and calendars.
Its agent can therefore receive a task, select tools and perform multiple operations rather than simply return text.
A request involving an upcoming appointment, for example, can require the system to inspect calendar information, reason about the event and interact with other connected resources. The model provides the decision-making layer while Odysseus exposes the tools through which those decisions can become actions.
··········
LOCAL EXECUTION CHANGES THE PRIVACY MODEL
The strongest architectural difference between Ajax and a conventional hosted assistant is where inference can occur.
A locally deployed Ajax instance can process prompts and agent state on hardware controlled by the user instead of transmitting each model request to a remote inference API.
This becomes particularly relevant when an agent works with email, documents, files, calendars, notes and persistent memory, because these systems can expose considerably more private information than an isolated chatbot conversation.
Local inference does not automatically make the complete system private.
An agent can still contact web services, external search systems, email servers or other APIs, and its privacy characteristics therefore depend on which integrations are enabled and how they are configured.
The important distinction is architectural: local model inference removes the requirement for a third-party model provider to process every inference request.
··········
AJAX HAS BEEN MODIFIED TO REDUCE MODEL REFUSALS
One of the most unusual elements of Ajax is its deliberate reduction of the refusal behavior present in the underlying model.
PewDiePie used Heretic, an open-source model-abliteration tool, as part of the process.
Abliteration attempts to identify and modify internal model behavior associated with refusals rather than simply adding a system prompt instructing the model to answer more requests.
This means the behavioral change occurs at the model level and remains present when Ajax is deployed locally.
The stated objective is not to remove every boundary. PewDiePie has described harm to oneself or others as a remaining limit.
Reducing refusals can make a locally controlled model more flexible, but it also transfers more responsibility to the surrounding application and the person operating it. This is particularly important for agents because their outputs can be connected to tools capable of modifying files, running shell commands or communicating with external services.
··········
PEWDIEPIE SAYS OPENAI BANNED HIM TWICE DURING DEVELOPMENT
The development process also produced an unusual dispute over model distillation.
PewDiePie says OpenAI suspended his access twice while he was developing Ajax. According to an account notice he displayed publicly, one suspension explicitly referenced distillation.
Distillation broadly describes using outputs from a stronger model to help train or improve another model. It has become an important issue as developers attempt to transfer capabilities from expensive frontier systems into smaller models.
The Ajax episode illustrates the tension between technically accessible model outputs and contractual restrictions governing how those outputs may be reused.
The suspensions are claims documented by PewDiePie from his own development process and should not be interpreted as independent evidence about the complete training dataset or as an assessment of whether any particular training method violated applicable terms.
··········
THE PROJECT FAVORS SPECIALIZATION OVER PARAMETER SCALE
Ajax represents a different development strategy from simply increasing parameter count.
A general-purpose frontier model needs to perform across a very broad distribution of tasks. A personal agent can instead be optimized around a smaller set of recurring operations and a known tool environment.
........
Approach | Large hosted frontier model | Ajax-style local agent |
Model scale | Very large | 9B parameters |
Inference | Cloud infrastructure | Local hardware |
Primary optimization | Broad general capability | Agent specialization |
Private data processing | Often requires remote inference | Can remain local |
Tool environment | Provider-controlled | Self-hosted workspace |
Model behavior | Provider policies | Greater operator control |
Compute cost | Recurring hosted inference | Local hardware resources |
Updates | Managed by provider | Managed by operator |
........
This does not mean a specialized 9B model becomes equivalent to a frontier model.
Instead, the relevant question becomes whether the smaller model is sufficiently capable for the specific tools and workflows it is expected to operate.
For routine agent tasks, that threshold can be substantially lower than the capability required to lead general-purpose model benchmarks.
··········
CONSUMER HARDWARE BECOMES PART OF THE AI AGENT STACK
Running Ajax locally moves model deployment from centralized AI infrastructure toward the user's own machine.
The practical hardware requirement depends heavily on quantization, context length, runtime and desired inference speed.
A 9B model represented at 16-bit precision requires roughly 18GB for raw model weights before runtime overhead.
Data Studios calculation: reducing the same 9 billion parameters to 8-bit representation lowers the theoretical raw weight footprint to approximately 9GB, while 4-bit representation reduces it to approximately 4.5GB.
Actual memory consumption is higher because inference also requires KV cache, runtime buffers, model metadata and the surrounding application.
The calculation nevertheless explains why the 9B scale is important: aggressive quantization can bring models of this size into the range of consumer hardware rather than requiring datacenter-class accelerators.
··········
LOCAL AGENTS ALSO CREATE A DIFFERENT SECURITY PROBLEM
Giving a local model access to private data solves only one part of the agent-security problem.
Odysseus can expose tools including files, shell access, web connectivity, email and persistent memory. An agent with those permissions can potentially perform consequential actions on the user's machine.
The security boundary therefore shifts.
With a conventional chatbot, the main concern is often what information is transmitted to a remote provider. With a local autonomous agent, permissions become equally important: which directories it can access, which commands it can execute, which accounts it can control and which external services it can contact.
Odysseus itself warns users to retain authentication, protect private information and avoid unnecessarily exposing model or service ports publicly.
Local-first architecture provides more direct control, but that control also makes the operator responsible for configuring the environment safely.
··········
AJAX IS AN EARLY RELEASE RATHER THAN A PROVEN FRONTIER COMPETITOR
Ajax should not yet be evaluated as if it were a mature frontier-model release.
At launch, there is not yet a sufficiently broad independent benchmark record to establish its performance against leading proprietary and open-weight models across reasoning, coding, tool use and long-horizon agent tasks.
PewDiePie has indicated that additional work is planned around reinforcement learning, quantization, benchmarking and further model refinement.
This makes the first Ajax release more significant as an architectural experiment than as a new benchmark leader.
The project asks whether a comparatively small model can become a useful everyday agent by being trained around the right tools, running continuously on personal hardware and operating within a self-hosted environment.
··········
AJAX SHOWS HOW LOCAL AI IS MOVING FROM CHATBOTS TO AGENTS
Local language models have traditionally been used primarily for private chat, experimentation and offline inference.
Ajax and Odysseus extend that model toward something more operational.
The model can sit inside an environment containing memory, files, research tools, email, calendars and system access. That transforms local inference from an isolated conversation into an agent architecture capable of interacting with the user's digital environment.
A 9B model will inevitably face capability limits that larger systems can overcome. But local agents do not necessarily need frontier performance on every task if specialization, tools and workflow design allow them to complete the narrower set of operations users perform repeatedly.
Ajax therefore represents a useful test of a broader direction in consumer AI: smaller specialized models running privately on personal hardware while the surrounding agent system supplies memory, tools and actions.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
