Where boundaries are made to be pushed.

The strongest open decision model we know of. Typed, calibrated decisions on long state, from a letter-slot readout, in one prefill pass.

JevBench hard (composite): 65.8
JevBench easy / original: 100.0 / 91.7
Judge-tier verification: 16/17
Long-state P50: 788ms

Deem 0.8B

The full Deem decision stack on CPU. Typed, calibrated choices, scores, and yes/no decisions served by our own Rust runtime. No GPU required.

Long-policy hold-out: 96.3%
Short-form latency: 362ms
Resident memory: 0.9GB
Exact-law zero-count: 1.000

DeepSeek-V4.1-Flash-1017 EXL3 2.9bpw

Our most downloaded repo this month: a 196.3 GiB EXL3 quant of 1017 that fits two GB10 boxes with 256K context. The Engram tables stay in the base repo, and the quality loss is 1.8% code and 0.4% prose.

Weight size: 196.3 GiB
Held-out code ppl: 1.9094
Held-out prose ppl: 6.4520
Context: 256K

DeepSeek-V4.1-Flash-1017 NVFP4

Preview

A 400 GiB NVFP4 transcode of 1017 in NVIDIA's ModelOpt layout. The expert weights are bit-exact; the Engram tables are lossy at cosine 0.9934. Not yet served by us.

Repo size: 399.5 GiB
Expert scales: Bit-exact
Engram cosine: 0.9934
Tensor layout: 188,245

DeepSeek-V4.1-Flash-1017 EXL3 3.5bpw

Preview

The higher-quality 1017 EXL3 build. 239.9 GiB, better held-out perplexity than 2.9bpw, and too large for a two-node GB10 pair, so it belongs on a three-node ring.

Weight size: 239.9 GiB
Held-out code ppl: 1.8895
Held-out prose ppl: 6.4341
Node budget: 3

MiMo-V2.6-Flash-MOPD NVFP4

Preview

The first W4A4 NVFP4 transcode of Xiaomi's 309B MOPD model. Experts are bit-exact at 169.9 GiB, and the activation scales are inherited from the RL sibling. Not served yet.

Weight size: 169.9 GiB
Expert values: Bit-exact
Index entries: 145,273
Activation scales: 94

Projects

Releases and research tracks

ShippedFeatured

Deem

An open-weight System One decision model family: typed, calibrated decisions (Choice, Score, Noul) over long unstructured state, read from a letter slot in one prefill pass. Shipped as deem-9b-v1 and a CPU-native deem-0.8-v1.

GLM-5.3-Flash on Blackwell

The first public 4-bit quantization of GLM-5.3-Flash, and the CUDA kernel that made a 320B multimodal model actually run on workstation Blackwell GPUs instead of only loading on them.

NVFP4 model releases

GGUF and safetensors builds of flagship open models with calibrated NVFP4 tensors, benchmarked honestly against stock K-quants on Blackwell hardware.

Consensus web search

A search API that queries Google, Bing and DuckDuckGo in parallel from Aleph Cloud's decentralized network and ranks results by cross-engine agreement.

LibertAI Code

A coding agent for the terminal and desktop that runs open models on confidential, TEE-backed inference, with no training on your code.

Experiment

Conscio

The consciousness layer for LLM agents: persistent memory, attention, drives, self-monitoring, reflection, and autonomous action in one inspectable runtime.

MCP server

A Model Context Protocol server built into the CLI that gives Claude Code, Cursor, Cline and any other MCP client access to consensus web search.

Experiment

Avatar

A browser avatar you can talk to, combining a 3D stage, browser speech APIs and LibertAI chat completions, with conversations kept in your browser.

GLM-5.3-Flash on vLLM, on GB10

A hand-written sparse-MLA CUDA kernel for sm_121 and a fix for vLLM's uninitialised NVFP4 MoE activation scale, which together take GLM-5.3-Flash from degenerate output to correct serving on a two-node DGX-Spark-class cluster.

Research

Post-quantum blockchain security review

An adversarial review of 30 live networks, migration efforts, standards proposals and on-chain verifiers, with high-priority findings moving through responsible disclosure.