AI startup LiquidAI, spun out of MIT, dropped two small models in one week — d1-3b (3 billion parameters) and d1-omni 600m (600 million parameters). Neither is a conventional large language model; both are pitched as "decision models." Our read of the community verdict: "something new, but not enough." What people were actually waiting for was a new generation of LFM (Liquid Foundation Model), and that didn't arrive.
What this is
LiquidAI was founded by former MIT researcher Ramin Hasani, and its core tech comes from his work on "liquid neural networks" — an architecture closer to biological neurons. The d1 series is the output of that line of thinking applied to decision tasks: small parameter counts, focused on control, automation, and Agent invocation (scenarios that need fast judgment), rather than writing or chatting. The company also released GGUF-format files on HuggingFace (a compressed format suited for local deployment), which means these models can run directly on ordinary computers and even edge devices (local terminals rather than the cloud).
Industry view
Supporters argue this is the right direction: Meta's Llama, Mistral, and other general-purpose large models are already plentiful — what the market truly lacks is small, specialized models that can run locally with low latency. The Agent era needs exactly this kind of "fast and accurate" tool. The criticism is more blunt — the Reddit discussion thread drew few comments, with users writing things like "No new LFM🙁," arguing the company failed to ship a flagship and is moving slower than rivals like Mistral and Alibaba's Tongyi Qianwen. Other analysts caution that inflated parameter claims and underwhelming real-world performance are recurring issues in the open-source community — a 3B-labeled model frequently loses to same-size competitors on benchmarks, so the proof will be in independent testing.
Impact on regular people
For enterprise IT: If d1 holds up in real-world testing, enterprises could deploy a local "decision small model" for risk control and process approval, no longer needing to call on hundred-billion-parameter general models for these tasks. Cloud costs could fall.
For working professionals: Limited relevance for now — decision models are primarily aimed at developer integration, and ordinary office scenarios aren't there yet. But developer community activity will be the leading indicator of whether this gains traction.
For consumer markets: No direct impact yet, but if "small models running locally" proves out, future phones, smart-home devices, and car infotainment systems could ship with stronger AI built in, without depending on the cloud.