Replicate snowflake/snowflake-arctic-instruct

Snowflake Arctic is a big, unusual open LLM - a sparse mixture-of-experts model with roughly 480B total parameters spread across 128 experts, but only about 17B active per token, released by the data-warehouse company Snowflake under Apache 2.0. It wasn't built to compete with GPT-4 on creative writing or general chat; Snowflake pitched it explicitly at enterprise tasks - SQL generation, coding, instruction-following for business use - and said as much when they released it. That's the model this node calls, hosted on Replicate.

Here's the honest heads-up before you drag this node into a graph: its input schema doesn't look like the other chat-LLM nodes in this pack. There's no prompt, no system_prompt, no sampling parameters. The only two optional fields are name and name_file, and neither is required - you could queue this node with nothing filled in at all.

How it works

Like every node in this pack, it's an API call - your inputs go to Replicate, Arctic runs on their infrastructure, and you get a STRING back. What's different here is figuring out what to actually put in those inputs, since the field names don't map onto the "prompt in, chat out" pattern you'd expect. A field / field_file pairing - one for typing text directly, one for pointing at a file with the same content - is a pattern that shows up on other Replicate model schemas, so name and name_file are plausibly that: type your prompt into name directly, or use name_file to point at a text file if your input is longer than you want to paste inline. Treat that as an educated read of the pattern rather than documented fact, though - this isn't a field-naming convention this pack's own README explains.

The inputs and outputs that matter

• name (optional, STRING) - the field to try putting your prompt or question into first.
• name_file (optional, STRING) - likely the file-input alternative to the above, for longer text.

Both are optional, and there's nothing else - no temperature, no max tokens, no system prompt. force_rerun is the only other input, for bypassing the cache. Output is a single STRING.

How to install it

One install covers every node in this pack.

cd ComfyUI/custom_nodes
git clone https://github.com/replicate/comfyui-replicate
cd comfyui-replicate
pip install -r requirements.txt

Restart, or via ComfyUI Manager: search ComfyUI-Replicate, install, restart. Set your token before launch:

export REPLICATE_API_TOKEN="r8_************"; python main.py

Token from replicate.com/account/api-tokens.

Common issues & troubleshooting

You queued it and got nothing useful back. Given the thin, oddly-named input surface, start by putting your actual question or instruction into name and see what comes back - if that doesn't work, check Replicate's own model page for snowflake/snowflake-arctic-instruct directly, since the node's field labels don't explain themselves and this pack's schemas are pulled once (via the maintainer's import_schemas.py) rather than kept continuously in sync with the model's own docs.

Node errors immediately. REPLICATE_API_TOKEN isn't set in ComfyUI's actual launch environment - the standard first-run failure across this whole pack. On portable Windows builds, a token set only as a global Windows environment variable has been reported not to get picked up; set it in the launch .bat instead.

You expected chat-style controls (temperature, max tokens) and they're not there. Correct, they're genuinely absent from this node's schema - unlike the Llama 2 nodes in this same pack, there's nothing to tune here beyond the two text fields.

Same output on identical input. Replicate's cache - flip force_rerun for a fresh call.