Open research from the LibertAI team
Model releases, benchmarks, agent experiments and infrastructure writeups, published as we build private, decentralized AI.
Projects
Some of these are in production behind api.libertai.io, and some exist to answer a question. The status badge tells you which is which.
NVFP4 model releases
shippedGGUF and safetensors builds of flagship open models with calibrated NVFP4 tensors, benchmarked honestly against stock K-quants on Blackwell hardware.
Read moreConsensus web search
shippedA search API that queries Google, Bing and DuckDuckGo in parallel from Aleph Cloud's decentralized network and ranks results by cross-engine agreement.
Read moreLibertAI Code
shippedA coding agent for the terminal and desktop that runs open models on confidential, TEE-backed inference, with no training on your code.
Read more
Conscio
researchAn autonomous agent runtime organized as a cognitive architecture, where attention, memory, goals and prediction are inspectable mechanisms rather than prompt roleplay, now with an eval harness built to falsify it.
Read moreMCP server
shippedA Model Context Protocol server built into the CLI that gives Claude Code, Cursor, Cline and any other MCP client access to consensus web search.
Read moreAvatar
experimentA browser avatar you can talk to, combining a 3D stage, browser speech APIs and LibertAI chat completions, with conversations kept in your browser.
Read moreOpen weights on Hugging Face
Quantized builds of the open models we serve, published under LibertAIDAI with benchmark numbers in every model card. Counts are from June 2026.
View on Hugging Face ↗- 9
- model repos
- 24,000+
- downloads
- 5
- base models
Notes
Writeups with methodology, the numbers that held up, and the ones that did not.
Does the cognitive architecture actually do anything? Measuring Conscio
We built Conscio around a real per-tick control loop, then measured it with a five-rung baseline ladder, six single-mechanism ablations and a self-report study on two models. Memory and reflection earn their keep, attention gating did not, and the most interesting number is about introspection.
NVFP4 GGUF performance on Blackwell: the consolidated numbers
Everything we have measured about NVFP4 GGUF serving across five models, three releases and one kernel generation, including the week the MoE result flipped in NVFP4's favor.
Ranking web search by cross-engine consensus
Our search API queries Google, Bing and DuckDuckGo in parallel from Aleph Cloud's decentralized network and ranks by agreement. Here is why consensus beats any single engine for agents, and what we never store.
Multi-token prediction on NVFP4 quants: a win, and a break-even
We added MTP draft-head variants of the Qwen3.6 NVFP4 builds. The dense 27B gains 22% generation throughput, the MoE gains nothing yet, and both model cards say so.
NVFP4 GGUF on Blackwell: what the benchmarks actually show
We shipped NVFP4 GGUF quants of three flagship multimodal models and benchmarked them against stock K-quants on an RTX 5090. Dense models win, MoE does not yet, and one number had to be re-run.