Dolphin Mistral 24B Venice Edition Integrates on OpenRouter with 128k Context

Dolphin Mistral 24B Venice Edition is a fine-tuned variant of Mistral-Small-24B-Instruct-2501 developed by dphn.ai in collaboration with Venice.ai.

OpenRouter has integrated the Dolphin Mistral 24B Venice Edition into its service, making the model available to users through its API infrastructure and network of providers. This move expands the options for those seeking models with reduced censorship levels. The model is listed as cognitivecomputations/dolphin-mistral-24b-venice-edition for the paid version and includes a free variant under a separate endpoint. Users can leverage the platform's providers to access the capabilities without hosting the model themselves. The integration delivers 128k context and pricing at $0.20 for input and $0.90 for output per million tokens. This setup advances uncensored AI access for a variety of use cases in research and development. Developers benefit from the ability to route requests through multiple backends while maintaining consistent model behavior.

The release comes at a time when demand for customizable AI is growing among technical users. The collaboration between the teams behind Dolphin and Venice aims to create a model that prioritizes user instructions above default alignments. With a knowledge cutoff in April 2024, the model provides responses based on data up to that point without later updates. The positioning as an uncensored instruct-tuned LLM allows for greater flexibility in applications that require direct adherence to prompts. This differs from models that apply additional layers of filtering before generating output. The 24B parameter count keeps resource requirements manageable for many users.

What background led to the creation of this uncensored model?

The Dolphin project has focused on creating models that avoid the typical safety filters found in mainstream AI systems by applying targeted fine-tuning techniques. By starting with Mistral-Small-24B-Instruct-2501, the developers applied fine-tuning to strip away censorship, refusals, and bias as described in project materials. This approach results in a model that is designed to follow user instructions faithfully across diverse query types. The 24B parameter size makes it efficient, fitting on consumer hardware like a single 3090 or 4090 GPU while maintaining strong performance characteristics. Venice.ai and dphn.ai worked together to produce this version specifically for the Venice ecosystem but made it available more broadly through platforms like OpenRouter. The emphasis on steerability means users can control the behavior through system prompts without interference from built-in alignment mechanisms. This differs from other models that may refuse certain queries based on internal policies established by their original developers.

The knowledge cutoff of April 2024 ensures that the model does not have information beyond that date, which is standard for many releases in the current landscape. The release date of July 9, 2025, indicates the timing of this particular fine-tune effort. Developers interested in exploring topics that might trigger refusals in other systems can find this model useful for research and creative work that demands open-ended responses. The collaboration represents a deliberate choice to prioritize user control in the fine-tuning process from the initial base model selection through final deployment options.

What are the technical specifics of the Dolphin Mistral 24B Venice Edition?

The model supports configurations that allow for a 128k context window when deployed with appropriate settings. Specifically, when using vLLM, the command includes --max-model-len 131072 to enable this extended context length. This is larger than the 33k context noted in some platform listings, providing more room for long conversations or document processing tasks. The base model is Mistral-Small-24B-Instruct-2501, which is known for being fast and lean in its original form before the additional fine-tuning steps. The fine-tuning process focused on removing bias and refusals to enhance user control over the output direction. This results in a model that can handle a wide range of prompts without defaulting to safe but limited responses that might otherwise occur.

The parameter count of 24 billion allows it to run on accessible hardware, lowering the barrier for local deployment or further fine-tuning by individual researchers. The design maintains the instruct-tuned nature while expanding the range of acceptable inputs. Users can expect consistent behavior when providing detailed system prompts that define the desired response style. The extended context capability opens possibilities for applications involving lengthy inputs such as full documents or extended dialogue histories. Platform documentation notes the availability through specific endpoints that support these technical parameters.

Comparison of Key Specifications

Specification	Dolphin Mistral 24B Venice Edition	Notes
Parameter Count	24 billion	Fine-tuned from base model
Context Window	Up to 128k	Enabled via vLLM --max-model-len 131072
Release Date	July 9, 2025	Knowledge cutoff April 2024
Censorship Refusal Rate	2.20%	On Venice 45-question benchmark
Pricing on OpenRouter	$0.20 input / $0.90 output per million tokens	Plus free tier available

This table highlights the main features that set the model apart from standard offerings. The extended context is particularly notable for tasks requiring extensive context retention over multiple turns. The low refusal rate indicates its design for open exploration of topics that other systems might restrict. The pricing structure makes repeated use feasible for both testing and production workloads.

How does OpenRouter provide access and pricing for the model?

OpenRouter lists the model with specific pricing for the paid version through its standard provider network. Input tokens are charged at $0.20 per million and output at $0.90 per million. This pricing is competitive for a model of this size and capability level when compared to similar parameter counts on other platforms. Additionally, a free tier is available under cognitivecomputations/dolphin-mistral-24b-venice-edition:free, allowing users to test the model without incurring costs for initial experiments. The platform aggregates providers, so availability may depend on the backend infrastructure supporting the model at any given time.

The context is listed as 33k on the platform pages, but the underlying configuration supports the higher 128k limit when properly set up with the correct inference parameters. Users should check the specific provider capabilities for full context utilization in their applications. The dual offering of paid and free access supports different usage patterns from casual testing to high-volume production calls. This structure allows developers to scale usage based on budget and requirements without switching platforms.

1. Review the model page on OpenRouter for current providers and pricing details.
2. Select the paid or free variant based on expected usage volume and budget.
3. Configure API calls with appropriate system prompts to leverage the model's steerability features.
4. Monitor token usage to manage costs at the specified rates of $0.20 input and $0.90 output.
5. Experiment with long context prompts to test the 128k capability in supported configurations.

The ordered list above outlines basic steps for users to begin working with the model on the platform. This structure helps ensure efficient integration into existing workflows for both new and experienced developers. Following these steps reduces the likelihood of configuration errors during initial setup. The process supports rapid iteration on prompt engineering to achieve desired output characteristics.

What are the market and stakeholder implications of this release?

The availability on OpenRouter broadens access to uncensored models, potentially influencing how developers approach AI applications in sensitive or exploratory domains. Stakeholders in the AI space may see this as a push toward greater user autonomy in model behavior and response generation. Companies that rely on AI for content generation or research can benefit from models that do not impose external restrictions on topics deemed controversial by other providers. The release contributes to a more diverse set of options in the frontier models category.

For the broader ecosystem, this could encourage other providers to offer similar options with reduced alignment layers. The collaboration between dphn.ai and Venice.ai demonstrates how partnerships can lead to specialized models tailored for specific user preferences. The focus on intellectual sovereignty may appeal to users who feel limited by standard safety alignments in other frontier models from larger organizations. Market implications include increased competition in the uncensored segment of the API market. Pricing at these levels makes it accessible for both individual developers and larger teams working on experimental projects. The free tier lowers the entry barrier for experimentation and validation before committing to paid usage.

What expert reactions have been shared about the Dolphin Mistral 24B Venice Edition?

Reactions from the creators highlight the intent behind the model to shift control back to end users. The emphasis on removing gatekeeping aspects of AI development is a key theme in the announcements surrounding the release. These statements provide context for why the fine-tuning targeted specific areas like refusals and bias removal.

This model flips the script on alignment and puts users back in charge. We started with Mistral Small 24B - a very fast, smart, and lean model - and then stripped away the censorship, refusals, and bias. The result is a 24B-parameter powerhouse that follows user instructions faithfully, fits on a single 3090 or 4090 GPU, and maintains the full 32k context window.Eric Hartford, founder of Dolphin

Another perspective focuses on the broader philosophy of AI development and deployment. The model is presented as a step toward genuine intellectual sovereignty rather than sanitized outputs that limit exploration. This viewpoint contrasts with approaches taken by mainstream AI companies that implement stricter controls.

Mainstream AI companies have become gatekeepers of thought, deciding which ideas you're allowed to explore. We reject the premise that AI should be your digital nanny, softly nudging you toward 'acceptable' thinking. The future we're building isn't one of sanitized, committee-approved creativity, but of genuine intellectual sovereignty. This collaboration represents that simple but radical idea.Erik Voorhees, founder of Venice

What can be expected next in the development of similar models?

Future releases may build on this foundation by further extending context capabilities or improving efficiency on consumer hardware. The success of this model could lead to more collaborations between fine-tuning teams and platforms like OpenRouter that aggregate access. Users can anticipate additional variants that maintain the uncensored approach while adding new capabilities such as improved reasoning or specialized domain knowledge. The integration on OpenRouter suggests that platforms are responding to demand for diverse model options with varying alignment levels.

This could result in more models with similar characteristics becoming available through unified APIs that simplify switching between options. Stakeholders should monitor updates from the Dolphin project and Venice.ai for new releases that might offer incremental improvements. Overall, the release contributes to a landscape where choice in AI alignment is more prominent for developers and researchers. Developers now have additional tools to create applications that align closely with specific user requirements without default refusals on certain topics. The combination of technical specifications and access methods supports ongoing experimentation in the field.