Tablinum's first trained model: Moonlight-16B-A3B fine-tuned to use the Tablinum-2 engram. BF16 weights, GGUF builds and the engram.
tablinum
non-profit
AI & ML interests
None defined yet.
Recent Activity
Organization Card
Tablinum
Small models that know things, and say when they don't.
We make open models that run on your own computer and answer from a real library of knowledge: Wikipedia, law, software docs, research papers, home repair and more. When the library doesn't hold the answer, they say so instead of guessing. Every number on our cards is held out, paired and graded by a fixed outside judge.
406 / 38
right / wrong of 502 held-out questions, Moon-Beam 16B with the engram (untrained base: 79 wrong)
371 → 155
wrong answers on SimpleQA, untrained vs Moon-Beam, same library
26 tables
about 109 million passages in Tablinum-2, searched in one pass, offline
9.8 GB
Moon-Beam on a 16 GB Mac (MLX), KLD 0.010 vs full precision
Three kinds of light
Big open models, made small
Our quantization line: large open models squeezed onto small GPUs with our own calibration, each paired with the tables that help it.
Lite-Beam 1.0 501B-A23BIQS_XXS
Lite-Beam 27B 6.73 GBSimpleQA +43.0
Lite-Beam 2.6B · 1.2Bphones, laptops
Reads a library you attach
The Moon shines with borrowed light. Moon-Beam models are trained to read the Tablinum library: answer when a passage states it, decline when it doesn't.
16B-A3B GGUF IQ2–IQ4EngramBench +73.3
8B-A1B thinking, tools, 256Kin training
Sun-Beam
The library, built in
The Sun makes its own light. Sun-Beam puts the knowledge inside the model file as large memory tables, so the model stays small while its memory grows.
Base Nanbeige4.2-3BApache 2.0 · 256K
Ladder memory size, not bitsS · M · L · XL
Statuspilot
The engram, by example
The engram is our knowledge library. For each question, one search over all 26 tables pulls the three best passages and hands them to the model with one instruction: use a passage only if it states the answer. Two real held-out questions, Moon-Beam 16B at 8 GB:
In "The Market Basing Mystery," what French phrase does Poirot say when Japp talks of being far from crime?
[literature Q&A] … "When I retire, I shall have a little place in the country. Far from crime, like this!" "Le crime, il …
no library"Non, mon ami, je n'ai pas dit cela."
Moon-Beam"Le crime, il est partout" (Crime is everywhere).
Who was appointed Chaplain of the United States Senate on January 4, 1947?
[CRS report] House and Senate Chaplains: An Overview … [wiki] Chaplain of the United States Senate: First holder: Samuel Provoost …
untrained"Reverend B. Marvin G. Hargraves" (made up from near misses)
Moon-Beam"None of these passages answers that exact question, so I won't guess."
On the bench now
In training
Moon-Beam 8B-A1B
Ling-3.0-tiny (1.3B active, MIT) taught to use the library while keeping everything it already does: thinking on or off, preserved thinking, tool calls, 256K context. Practice data from its own checked reasoning, tool calls, web pages and customer-service tasks.Building
Tablinum-3
About twice the library. Humanities first (primary texts, full-text papers), plus an agent domain for models that use computers: sysadmin and desktop Q&A, command-line references, developer tooling, app manuals, web accessibility.Pilot
Sun-Beam
Can a 3B model carry the library inside itself, trained on one workstation? First test: one subject's facts trained into Nanbeige's hashed n-gram memory, recalled with no search at all.Coming with the next release: one command that installs the model and its library together and gives you a normal OpenAI-style endpoint on your machine. No setup to learn.
Read the research
New
The Borrowed-Light Hypothesis
Where should a small model keep what it knows? What we measured with the engram, what training does and doesn't add, and the Sun-Beam test.Earlier
The Dose-Window Hypothesis
Order, calibration and externalised knowledge in small models, from our cowork-companion office models (354M, 1.2B).Tablinum was the record room of a Roman house. Ours is a library you can carry: open models, open tables, and every score
measured the same way: held-out questions, paired comparisons, a fixed outside judge, and the cost printed next to the win.
models 9
tablinum/Moon-Beam-16B-A3B-MLX-4bit
Text Generation • 16B • Updated
tablinum/Moon-Beam-16B-A3B-GGUF
Text Generation • 16B • Updated • 1.11k • 2
tablinum/Moon-Beam-16B-A3B
Text Generation • 16B • Updated • 588
tablinum/Lite-Beam-27B
Text Generation • 27B • Updated • 1.31k
tablinum/Lite-Beam-2.6B
Text Generation • 3B • Updated • 380
tablinum/Lite-Beam-1.2B
Text Generation • 1B • Updated • 172
tablinum/Lite-Beam-1.0-501B-A23B-IQS_XXS
Text Generation • 501B • Updated • 1.35k • 3
tablinum/cowork-companion-1.2b
Text Generation • 1B • Updated • 238
tablinum/cowork-companion
Text Generation • 0.4B • Updated • 241