Eric Hartford's Cognitive Computations team trained this model in 3 days on 4x A100 GPUs using QLoRA, fine-tuning Mistral's Mixtral 8x7B mixture-of-experts architecture with ~46.7B total parameters and ~12.9B active per forward pass. The training data was systematically filtered to remove alignment refusals, producing a deliberately uncensored model for code generation, creative writing, and open-ended conversation.
Discover the power and diversity of large language models available with Telnyx. Explore the options below to find the perfect model for your project.
| Organization | Model Name | Tasks | Languages Supported | Context Length | Parameters | Model Tier | License |
|---|---|---|---|---|---|---|---|
| No data available at this time, please try again later. |
Powered by our own GPU infrastructure, select a large language model, add a prompt, and chat away. For unlimited chats, sign up for a free account on our Mission Control Portal here.
Dolphin 2.5 Mixtral 8x7B is an open-source model created by Eric Hartford, built on Mistral AI's Mixtral 8x7B mixture-of-experts architecture. It was trained using qLoRA on the OpenHermes dataset for general-purpose text generation, coding, and conversational tasks.
Dolphin 2.5 Mixtral is designed as an uncensored model, meaning it does not include built-in content filtering or refusal mechanisms. This makes it highly compliant with user instructions but also requires responsible deployment practices since it will not refuse harmful requests on its own.
Dolphin 2.5 Mixtral 8x7B scores approximately 70.8% on MMLU (community benchmarks), in line with the base Mixtral 8x7B Instruct (70.6%) on the same sheet. The uncensored fine-tuning preserves the base model's benchmark quality while removing safety guardrails. Its Arena ELO of 1,063 places it below the base Mixtral Instruct (1,114), likely reflecting the removal of alignment training.
The cost of running Dolphin 2.5 Mixtral 8x7B with Telnyx Inference is $0.0003 per 1,000 tokens. Processing 1,000,000 uncensored generation tasks at 1,000 tokens each would cost $300, the same as the base Mixtral 8x7B Instruct on the sheet but without safety guardrails.
Dolphin models do not have built-in safety guardrails or moderation layers. Users who deploy Dolphin models are responsible for implementing their own safety measures and content filtering through system prompts, output filters, or external moderation tools.
Dolphin 2.5 Mixtral excels at general text generation, creative writing, code generation, and open-ended conversation. Its mixture-of-experts architecture provides strong performance comparable to GPT-3.5 Turbo on many benchmarks while being fully open-source and self-hostable.
