What is Dolly v2 3B?
Dolly v2 3B is an open-weight causal language model from Databricks. A causal language model generates text one token at a time, using the preceding text to predict what should come next. In practical terms, it accepts an English instruction or question and produces a text response.
The model contains approximately 2.8 billion parameters and is based on EleutherAI's Pythia-2.8B foundation model. Databricks fine-tuned it on approximately 15,000 instruction-response examples called databricks-dolly-15k. Those examples cover brainstorming, classification, closed-question answering, generation, information extraction, open-question answering, and summarization.
Dolly v2 3B is best understood as a downloadable model checkpoint rather than a hosted chatbot or current general-purpose API product. Its weights are available through the databricks/dolly-v2-3b repository, allowing developers to run, inspect, quantize, customize, or serve the model using their own infrastructure.
Where it fits in Databricks' model lineup
Dolly v2 3B belongs to Databricks' Dolly family of open models and is now a legacy model. Its significance is primarily its accessibility: it provides a relatively small instruction-tuned checkpoint that can be used without depending on a Databricks-hosted inference endpoint.
It should not be evaluated as a current frontier model. The supplied research identifies no first-party hosted API for this checkpoint, no web-search integration, and no provider-operated token billing. Developers choosing Dolly v2 3B are therefore choosing local control and low infrastructure scale over the convenience, current knowledge, managed reliability, and advanced capabilities normally associated with newer hosted models.
Technical specifications
The published configuration identifies Dolly v2 3B as a GPT-NeoX causal language model. Its 2,048-token maximum position length limits how much prompt and conversation history can be processed in one context. This is sufficient for short instructions and small documents, but restrictive for long reports, extended conversations, large retrieval results, and codebases.
| Specification | Verified detail |
|---|---|
| Provider | Databricks |
| Model identifier | databricks/dolly-v2-3b |
| Release | April 2023 |
| Model status | Legacy open-weight model for downloadable, self-hosted inference |
| Architecture | GPTNeoXForCausalLM |
| Parameter count | Approximately 2.8 billion |
| Base model | EleutherAI Pythia-2.8B |
| Fine-tuning data | Approximately 15,000 instruction-response records |
| Context length | 2,048 tokens |
| Input | English text |
| Output | Generated text |
| Maximum output tokens | Not specified in the supplied research |
The original checkpoint configuration specifies bfloat16 weights. Actual memory use and generation performance depend on the serving software, hardware, precision, and any quantization applied by the operator.
Capabilities and strengths
Dolly v2 3B can follow straightforward natural-language instructions and generate responses for lightweight language tasks. Suitable examples include asking it to summarize a short passage, classify text into a small set of categories, extract named details, brainstorm ideas, answer a simple question, or produce a short piece of text in a requested format.
Its main practical strength is deployability. Because the checkpoint can be downloaded, developers are not required to send prompts to a Databricks service or pay a provider for every generated token. The relatively small model size also makes it more approachable for local experimentation than substantially larger models, especially when using reduced precision or quantization.
Open weights can be valuable when a team needs to study model behavior, test prompting methods, build an educational project, or investigate customization. Fine-tuning is possible because the weights are available, although the research does not identify a Databricks-operated fine-tuning service for Dolly v2 3B.
Reasoning, coding, and quality limitations
Databricks does not describe Dolly v2 3B as state of the art. Its model documentation reports problems with syntactically complex prompts, programming, mathematical operations, factual accuracy, dates and times, open-ended questions, hallucination, exact-length list generation, stylistic imitation, humor, and some forms of letter writing.
These limitations affect how the model should be used. It may produce a plausible-sounding answer even when it lacks the information needed to answer correctly. It should not be treated as an authoritative source for current events, dates, calculations, or specialized factual research. The supplied research does not establish an authoritative knowledge-cutoff date for the exact checkpoint; the dates associated with instruction-data generation should not be interpreted as the model's complete pretraining cutoff.
Reasoning and coding ratings in the supplied model data are editorial estimates rather than Databricks benchmarks. They assign reasoning a score of 3 out of 10 and coding a score of 2 out of 10. These ratings are consistent with using the model for simple transformations and basic instruction following, but not for dependable multi-step reasoning, software development, or mathematical work.
Modalities, tools, and API support
Dolly v2 3B supports text input and text output only. It does not natively accept images, audio, or video, and it does not produce image, audio, video, music, embedding, or speech output. The model is not documented as having built-in web search, function calling, or tool-use support.
Local serving frameworks may provide conveniences such as streaming generated text, but that is a property of compatible inference tooling rather than a Databricks-hosted API feature. The supplied data records streaming as available through local generation tooling while also noting that no provider-hosted streaming API is documented.
Speed, cost, and pricing
There is no official per-token input or output price for the downloadable Dolly v2 3B checkpoint. Instead, the operator pays for the resources used to run it, such as GPU or CPU time, storage, electricity, and serving infrastructure. The final cost depends on hardware, precision, quantization, traffic, and the chosen inference stack.
The model's editorial speed score is 7 out of 10 and its editorial cost score is 9 out of 10. These are subjective comparative estimates, not provider-published measurements. They reflect the likely advantage of a smaller, self-hosted model over larger models in resource requirements, but they do not guarantee a particular tokens-per-second rate or total operating cost.
Local deployment can be inexpensive for intermittent experiments, but a hosted model may become more economical when a team would otherwise need to maintain reliable hardware, monitoring, scaling, security, and availability. Dolly v2 3B therefore has a low direct model price, but not necessarily zero total cost of ownership.
Best use cases
- Learning how instruction-tuned language models work.
- Local inference experiments on modest hardware.
- Prompt-engineering and model-behavior research.
- Short English summarization, classification, extraction, and brainstorming tasks.
- Projects that require downloadable weights or offline-style self-hosted operation.
- Exploring quantization, fine-tuning, and custom serving pipelines.
For these uses, the ability to control the checkpoint can matter more than achieving the strongest possible answer quality. The model is especially suitable when the task is short, the language is English, and a human or downstream process can review the output.
When to choose Dolly v2 3B
Choose Dolly v2 3B when open weights, local execution, educational value, or infrastructure control are central requirements. It can be a reasonable starting point for a small proof of concept where the model's responses are reviewed and the application does not require current information.
A newer hosted model is generally more appropriate when the application needs strong reasoning, reliable coding, mathematics, long-context processing, current knowledge, managed scaling, web search, tool calling, or production-grade safety controls. A larger or more recent open-weight model may also be preferable when self-hosting is required but Dolly v2 3B's quality limitations are unacceptable. The supplied research does not identify a specific successor or provide benchmark comparisons, so such a choice should be validated against the application's own prompts and evaluation set.
Deployment and licensing considerations
The model can be loaded locally with the Transformers library using the custom instruction-generation pipeline described in its repository. The original examples require trusting repository-provided remote code and recommend reduced-precision loading where supported to reduce memory requirements.
Databricks describes Dolly v2 3B as commercially usable under the license published with its model repository. Before redistribution or commercial deployment, review the exact repository license, model terms, dependency licenses, and any obligations associated with the intended use. Because the model is self-hosted, the deploying organization is also responsible for access controls, monitoring, output review, and operational reliability.
Dolly v2 3B remains useful as an accessible example of an instruction-tuned open-weight model, but its role is specialized. It is strongest as a compact, controllable model for local experimentation and lightweight text generation—not as a replacement for modern, managed, long-context, or frontier-quality systems.