ACS Research opens free platform serving large open base models (Llama 3.1 405B, Trinity-Large-TrueBase) to LLM psychology researchers

ACS Research has opened application-based free access to a hosted platform, ACS Infra, after a summer pilot. It serves Llama 3.1 405B, Llama 3.1 8B and Arcee's Trinity-Large-TrueBase, a raw 398B-parameter pre-training checkpoint. Researchers get a completions API with logprobs, activation harvesting and steering, a browser workbench and a community Discord. Capacity is sized for about 70 to 100 active researchers, funding from Coefficient Giving is meant to cover at least a year, and applications are reviewed within a few days with no publications or affiliation required.

The people affected are independent and small-team researchers studying how post-training changes model behaviour. Base models at this size are rarely offered through APIs, and self-hosting them is costly, so this could let them check whether findings from small models hold at 400B+ scale. The effect is narrow. It is a few dozen users, no new models are released, and the access depends on one funder and one operator. It is a useful resource for a niche, not a shift in who controls capable models.

The framing that this "lowers the barrier" is true but modest, given the 70 to 100 researcher cap and the application gate. "At least a year" is a funding intention, not a commitment, and the article gives no usage limits, rate caps or compute quotas. The source is the operator's own post on LessWrong, so claims about demand and usefulness are untested by anyone else.

1. 01Whether the platform publishes how many researchers are admitted and how long the waitlist or review time actually is, against the stated 70 to 100 capacity and few-day turnaround.
2. 02Papers or posts over the next 6 to 12 months that cite ACS Infra for results at 400B scale, which would show whether it produces research that would not otherwise exist.
3. 03Any change to funding, usage limits or model lineup (for example newer open base models added or Llama 3.1 dropped) as the one-year mark approaches in autumn 2027.