Snowflake open-sources the 480B hybrid model Arctic

What was announced

On 24 April 2024, Snowflake announced Arctic, a large language model it presented as built for enterprise AI rather than for general chat ^[1]. The framing mattered: instead of leading with benchmark scores or consumer use cases, the announcement organised the model around the kinds of tasks that show up inside companies — writing and correcting database queries, producing code, and following multi-step instructions ^[1].

Those three categories are worth separating. SQL generation means turning a question asked in ordinary language into a query that a database can run. Coding means writing, completing or repairing program text. Instruction following means carrying out a request that has several parts or constraints, rather than answering a single question. A model that is good at these is useful to a business in a fairly direct way, because the output can be checked by running it — a query either returns rows or fails, code either compiles or does not.

What shipped alongside the weights

Arctic was not released as a bare set of parameters. Together with the weights, Snowflake published a cookbook of training recipes and research notes ^[1].

The distinction is worth spelling out. The weights are the numbers a model learns during training; loading them lets someone run the model, but it does not by itself explain how the model was produced. Training recipes describe the procedure — how data was prepared and combined, how training was scheduled — and research notes record what the team observed along the way. In the announcement, Snowflake argued that openness should include data recipes and not just weights ^[1]. That is a stronger claim than simply posting a checkpoint, because a model's behaviour depends heavily on what it was trained on, and weights alone do not disclose that.

Where it was available

Distribution was broad from the start. The model shipped on Hugging Face, the NVIDIA API catalog and Replicate on the day of announcement ^[1]. Further access was promised through Snowflake Cortex, AWS, Azure, Lamini, Perplexity and Together ^[1].

Putting a model on several platforms at once lowers the cost of trying it. Hugging Face is where many practitioners go to download weights; an API catalog lets developers call a model without hosting it; other platforms serve the same model to their own customers. For a company selling into enterprises, meeting customers inside the tools they already use is often as important as the model itself.

Who supplied the plumbing

Two partners were credited with the infrastructure. AWS helped build the training cluster, and NVIDIA enabled serving through TensorRT-LLM ^[1].

Training a large language model requires a cluster of accelerators working together for a sustained period, and assembling that cluster is an engineering project in its own right. Serving is a different problem: once the model exists, it has to answer requests quickly and cheaply enough to be useful, which is what an inference optimisation library such as TensorRT-LLM is for. Splitting the work this way — one partner for building the model, another for running it — is common for models that are meant to be deployed at commercial scale rather than studied in a lab.

The Cortex offer

Snowflake also attached a time-limited offer to the release: Cortex customers got free access to Arctic until 3 June 2024 ^[1]. Cortex is Snowflake's own route for bringing language models into its data platform, so the offer put the model in front of exactly the customers the announcement was addressed to, with a deadline attached.

Why it mattered

The release is best read as a statement about where language models were heading commercially. Arctic was positioned as an enterprise model, distributed through many channels on day one, backed by named infrastructure partners, and accompanied by an argument that releasing weights without describing the data behind them is incomplete openness ^[1]. Each of those choices is a position rather than a technical detail: enterprise first, wide distribution, and a broader definition of what it means to open a model.