A / THE MODEL
Start at
the weights.
ABLE-Origin explores training a language model from scratch: architecture, tokenizer, training, and the evidence left by a run.
Read the model record ↗ABL—000 / THE WORK, MADE VISIBLE
Inside
the lab.
Follow the work.
An independent AI lab exploring local models, agents, persistent memory, and the engineering between an idea and something you can inspect.
Enter the laboratory ↓01 / WEIGHTS
02 / ROUTES
03 / EVIDENCE
FIG. 00RESEARCH FIELD
SCHEMATIC / NOT TELEMETRY
FROM MODEL WEIGHTS TO PUBLIC RECORDSCROLL TO DESCEND ↓
§01 / THROUGH THE CROSS-SECTIONGo beneath
Go beneath
the interface.
EXPLAINER / CONCEPTUAL CROSS-SECTION
Start at
B / THE SYSTEM
Make the
parts meet.
Local inference, agent routes, memory and application data controls form different engineering problems. A system diagram should make their boundaries easier to understand.
Inspect the public source excerpt ↗C / THE EVIDENCE
Leave a trace
you can question.
Measurements become useful when the configuration, source and limits stay attached. Follow the evidence into a real experiment below.
Inspect an inference result ↓D / THE PUBLIC EDGE
A public record
has a boundary.
This proposal separates public project evidence from owner decisions and private runtime data. The journal below contains no fabricated updates.
Approach the journal ↓§02 / SELECTED RESEARCH RECORDS
The result.
The route to it.
Inspectable work, its source, and the limits that travel with it.
REC / 01MODEL DEVELOPMENTSOURCE-REPORTED · HISTORICAL
ABLE—
Origin.
A language model, starting from initial weights.
29.9 million parameters. Custom architecture and tokenizer work. Trained on a consumer GPU with TinyStories.
Inspect source ↗FIG. 01 / SIX REPORTED VALIDATION CHECKPOINTS
Best reported: 1.341 at 60,000 steps
Partial checkpoints, not a full training curve.
August 2026 source report. Dashed joins guide the eye. Raw logs and independent reproduction are not bundled here.
Article and context ↗Inspect the six data points
| Training step | Loss | Note |
|---|---|---|
| 500 | 2.379 | Reported |
| 1,000 | 1.964 | Reported |
| 1,500 | 1.837 | Reported |
| 42,000 | 1.374 | Reported |
| 60,000 | 1.341 | Best reported |
| 61,894 | 1.344 | Final reported |
REC / 02LOCAL INFERENCESOURCE-REPORTED · AUG 2026
27B.
8 GB.
A constraint becomes the experiment.
Quantization, offloading and decoding under a tight GPU memory budget. Different configurations, visible tradeoffs.
Read the experiment ↗27B + 4B DRAFTFailed to load
No throughput measurement. A failure is not a zero.
RTX 3070 Ti 8 GB · i7-12700K · WSL2. Source-reported outcomes, not rerun for this atlas. Prompt counts, repetitions and timing details are incomplete; comparisons are illustrative.
Inspect source and limitations ↗REC / 03LOCAL AGENT SYSTEMSPUBLIC SOURCE EXCERPT
Connect.
Keep boundaries.
Explorations in local model routes, persistent memory and application-level data controls.
Read about ABL Core ↗EXPLANATORY SCHEMATIC / NOT LIVE TOPOLOGY
REQUEST→ROUTE→MODEL
+ MEMORY
+ MEMORY
CONTEXT / DATA CONTROLS↓EVIDENCE→HUMAN
REVIEW
REVIEW
§03 / THE INSTRUMENT
Different layers.
One inquiry.
A spectral index of the work. Select a layer to inspect its role.
01 / MODEL
From initial weights
to a language model.
Architecture, tokenizer and training work in ABLE-Origin.
Read the model experiment ↗02 / SERVING
The memory budget
is part of the design.
Quantization, GPU offload and decoding experiments under hardware constraints.
Read the local inference article ↗03 / SYSTEM
Useful boundaries
are visible boundaries.
Public source excerpts explore agent routes, memory and application data controls.
Inspect ABL Core ↗04 / RECORD
A result needs
its surrounding facts.
Source, timing, failures, limitations and review belong beside the conclusion.
Read the latest lab notes →INDEX OF PUBLIC WORKNO SERVICE HEALTH IS IMPLIED
§04 / THE PUBLIC JOURNAL
Behind every entry,
a door to cross.
Real work. Inspectable evidence. A considered public record.
LATEST ENGINEERING NOTES
Containment. Rejections. Boundaries.
A development log should carry its source, work date, review and limitations. Read the November 4 and 5 candidate results: owned process cleanup, Promise rejection regressions and service-identity fences.
Open the journal →OWNER DECISIONS AND PRIVATE RUNTIME DATA STAY OUTSIDE THIS PUBLIC DESIGN.ABSENT BORN LABS / Asif
The experiment
continues.
An interactive atlas of the lab’s engineering work. Public writing and source links open the current published work. The drawings describe architecture. Dated lab notes carry the measured results.