Weightsmithregistry

acme / atlas-70b

atlas-70b

A 70-billion-parameter text model, Apache-2.0, with a full training-data manifest and reproducible evaluation.

pull
ws pull acme/atlas-70b@v3

# or a specific shard set
ws pull acme/atlas-70b@v3 --format safetensors --shards 00001-00038

# quantised variant
ws pull acme/atlas-70b-gguf-q4@v3

Artefact

Version
v3 (2026-07-18)
Digest
sha256:8f21c0a2b4…
Format
safetensors, 38 shards
Size
142.4 GB
Parameters
70.6 B
Context
131,072 tokens

Provenance

Base model
trained from scratch
Training tokens
14.2 T
Data manifest
published, 41 sources
Compute
4,096 × H100, 38 days
Signature
verified — acme publishing key
SBOM
attached

Licence and use

Licence
Apache-2.0
Commercial use
permitted
Redistribution
permitted with notice
Derivatives
permitted
Attribution
required in documentation
Export note
none

Evaluation

Run with the published harness; every number is reproducible from the recorded configuration.

BenchmarkScorePrevious versionHarness
Reasoning suite74.271.8v2026.6
Code generation68.964.1v2026.6
Long-context recall91.488.0v2026.6
Multilingual QA70.169.4v2026.6
Instruction following82.679.2v2026.6
Refusal calibration88.084.7v2026.6

Files

model-00001-of-00038.safetensorssha256:4f2a91bd…3.9 GB
model-00002-of-00038.safetensorssha256:8c31f0a2…3.9 GB
model.safetensors.index.jsonshard map412 KB
tokenizer.jsonbyte-level BPE18 MB
config.jsonarchitecture2.1 KB
provenance.sigdetached signature566 B