Dolphin 2.9.1 Yi 1.5 34B
dphn/dolphin-2.9.1-yi-1.5-34b
published May 2024 · updated Sep 2025
Dolphin 2.9.1 Yi 1.5 34B is a text-generation model that is a full-parameter fine-tune of Yi-1.5-34B with instruction, conversational, coding, and agentic skills.
specs
| Task | Text Generation |
| Architecture | LlamaForCausalLM (based on Yi-1.5-34B) |
| Parameters | 34.4B |
| Context Length | 8k (trained with 8k, base 4k) |
| License | Apache 2.0 |
about this model
Dolphin 2.9.1 Yi 1.5 34b is a text-generation model based on Yi-1.5-34B, fine-tuned with full parameter updates (FFT) at 16-bit precision. It achieves 77.4 on the MMLU benchmark at the 34B parameter scale. The model uses the ChatML prompt format and is designed for instruction following, conversational interaction, coding tasks, and initial agentic capabilities including function calling.
Training and Architecture
Trained on a mixture of datasets including Dolphin-2.9, OpenHermes-2.5, CodeFeedback, Orca-Math, agent-instruct, toolbench, and dolphin-coder data, the model underwent three epochs with a learning rate of 1e-5 and a total batch size of 64. It was fine-tuned with a sequence length of 8k using a rope theta of 1,000,000.0, while the base architecture supports a maximum positional embedding of 4k. The model contains approximately 34.4 billion parameters.
Performance
Benchmark results are summarized below:
| Benchmark | Score |
|---|---|
| MMLU | 77.4 |
Capabilities
- Instruction following and multi-turn conversation.
- Code generation and translation.
- Function calling and tool use (React, ToolBench).
- Uncensored behavior: training data filtered to remove alignment and bias; users should implement their own safeguards before exposing as a service.
Usage Notes
The model is distributed under the Apache 2.0 license. It is hosted as a managed API by Gigarouter, supporting OpenAI-compatible endpoints for text generation inference. The model is tagged for text-generation-inference and Inference Endpoints.
best for
- ·Conversational AI assistants and chatbots
- ·Code generation, translation, and debugging
- ·Function calling and agentic tool-use workflows
FAQ
It excels at instruction following, conversation, coding tasks, and agentic capabilities (function calling).
It achieves 77.4 MMLU, one of the highest scores among 34B models, and is fully fine-tuned on diverse datasets.
It is licensed under Apache 2.0, allowing commercial use.
It uses the ChatML format with <|im_start|>system, user, and assistant tokens.
Use the gigarouter OpenAI-compatible endpoint with your API key to send ChatML-formatted messages.
We're benchmarking and onboarding Dolphin 2.9.1 Yi 1.5 34B as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.