dolphin-2.9.1-yi-1.5-34b
Dolphin 2.9.1 Yi 1.5 34b is a 34B-parameter instruction-tuned LLM based on Yi-1.5-34B, fine-tuned on diverse conversational, coding, and agentic datasets. It reports 77.4 MMLU performance, uses ChatML formatting, and is positioned as an uncensored model with minimal alignment filtering. Licensed under Apache 2.0, it is ungated and suitable for self-hosted or custom application deployment.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | dphn |
| Parameters | 34.4B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 4.6M |
| Likes | 65 |
| Last updated | 2025-09-08 |
| Source | dphn/dolphin-2.9.1-yi-1.5-34b |
What dolphin-2.9.1-yi-1.5-34b is
Full-weight fine-tuned derivative of 01-ai/Yi-1.5-34B using Axolotl on 8k sequence length (base model max 4k). Trained with RoPE theta 1000000.0 to extend context. Training involved 3 epochs over curated multi-source datasets (conversational, coding, math, agentic function-calling) using 8×H100 GPUs with DeepSpeed ZeRO-3, BF16 precision, AdamW optimizer. Validation loss: 0.4425. Prompt template: ChatML. No LoRA; full-parameter update.
Run dolphin-2.9.1-yi-1.5-34b locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="dphn/dolphin-2.9.1-yi-1.5-34b")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
ESTIMATE: 34B fp16/bf16 ≈ 68 GB VRAM (single GPU). A100 (80GB), H100 (80GB), or multiple GPUs recommended. Quantization (4-bit ≈ 17 GB, 8-bit ≈ 34 GB) feasible but not benchmarked in card. Card confirms training used 8×H100.
Model is full-weight fine-tuned from Yi-1.5-34B; no LoRA layers used. Further fine-tuning is possible via standard frameworks (Axolotl, LitGPT, Hugging Face Trainer) but requires significant VRAM or distributed training. QLoRA adaptation likely feasible on single A100/H100 but not validated in documentation.
When to avoid it — and what to weigh
- Production Safety-Critical Systems — Model is explicitly uncensored and designed to comply with any request, including unethical ones. Card advises implementing your own alignment layer. Not suitable for high-stakes applications (healthcare, legal, finance) without additional safety review and filtering.
- Long-Context Tasks Beyond ~8k Tokens — Base model max context is 4k; extended to 8k via fine-tuning rope theta. Longer sequences may degrade performance. Use alternative models if 32k+ context is required.
- Edge Deployment or Constrained Hardware — 34B parameter count requires significant VRAM. Quantization (4-bit, 8-bit) possible but not described in card. Not practical for mobile, IoT, or very resource-limited environments without post-hoc optimization.
- Scenarios Requiring Model Transparency/Auditability — Trained on GPT-4 generated data and multiple closed-source sources. Full training data lineage and reproducibility are not publicly documented. Not ideal if provenance is a compliance requirement.
License & commercial use
Apache License 2.0. Permissive OSI-compliant open-source license granting broad usage rights including modification and redistribution.
Apache 2.0 explicitly permits commercial use. Card states: 'We grant permission for any use, including commercial.' No restrictions noted. However, model is trained on GPT-4 generated data (unclear if OpenAI terms permit derivative commercial models); buyer should review OpenAI terms independently. Base model (Yi-1.5-34B) license should also be verified.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Adequate |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Good |
| Assessment confidence | High |
Model is uncensored and designed for minimal refusal behavior. This creates risk of generating harmful, illegal, or misleading content if exposed without filtering layer. Card advises implementing alignment safeguards before production deployment. No vulnerability assessment, adversarial robustness evaluation, or prompt-injection defense analysis provided. Buyer responsible for downstream content filtering and use-case constraints.
Alternatives to consider
Meta Llama 2 34B / Llama 3 70B
Larger community, stronger alignment out-of-the-box, comparable performance. Trade-off: Llama 2 is older; Llama 3 70B has higher VRAM footprint but better safety.
Mistral 7B / Mixtral 8×7B
Lower VRAM requirements, faster inference, strong code/reasoning performance. Trade-off: Smaller capacity; may underperform on complex reasoning vs. 34B.
Yi-1.5-34B (base model)
Smaller footprint if untuned performance suffices; avoid Dolphin's uncensored-model liability if alignment is required. Trade-off: No instruction tuning; weaker on chat/agent tasks.
Ship dolphin-2.9.1-yi-1.5-34b with senior software developers
Start with quantized inference on a single 40GB+ GPU using vLLM or Ollama. Verify alignment safeguards for your use case before production exposure. Contact our team to discuss private-hosted or custom integration options.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
dolphin-2.9.1-yi-1.5-34b FAQ
Can I use Dolphin 2.9.1 in a commercial product?
What GPU do I need to run this model?
Is this model safe to expose to end users?
How does context length compare to other 34B models?
Work with a software development agency
DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If dolphin-2.9.1-yi-1.5-34b is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.
Ready to Deploy Dolphin 2.9.1?
Start with quantized inference on a single 40GB+ GPU using vLLM or Ollama. Verify alignment safeguards for your use case before production exposure. Contact our team to discuss private-hosted or custom integration options.