Barcelona Supercomputing Center

weights

open(Apache-2.0, safetensors on HF BSC-LT/ALIA-40b)

data

components-listed(60+ named sources described in the model card, but the pretraining corpus "will not be released or distributed to third parties")

code

open(all training scripts + configuration files at github.com/langtech-bsc/alia)

checkpoints

open(partially pretrained revisions tagged v0.5 and v0.9 on HF)

license

Apache-2.0(OSI)

ALIA-40B's Apache-2.0 weights ship with the complete training code and recipe on GitHub, but the pretraining corpus is documented rather than released or reconstructable. That gap keeps it at open weights rather than the open-source tier that OLMo, Pythia and Apertus reach, the last by publishing data-reconstruction scripts.

• https://api.github.com/repos/langtech-bsc/alia recorded 2026-08-13

  `private: false`, `archived: false`, license `Apache-2.0`. The training scripts and configuration files the model card points at are published.

• https://cdn.jsdelivr.net/gh/langtech-bsc/alia@main/ recorded 2026-09-29

  jsDelivr mirror of the GitHub repository langtech-bsc/alia at main lists the directories "/gh/langtech-bsc/alia@main/configs/", "/gh/langtech-bsc/alia@main/launchers/" and "/gh/langtech-bsc/alia@main/tokenizer/" plus a LICENSE file, so the training code tree is public.

• https://cdn.jsdelivr.net/gh/langtech-bsc/alia@main/configs/ recorded 2026-09-29

  jsDelivr listing of the repository configs directory shows "/gh/langtech-bsc/alia@main/configs/bsc_40b.yaml", "/gh/langtech-bsc/alia@main/configs/bsc_40b_32k.yaml" and "/gh/langtech-bsc/alia@main/configs/40B_quality_data.yaml".

• https://cdn.jsdelivr.net/gh/langtech-bsc/alia@main/LICENSE recorded 2026-09-29

  Repository LICENSE file (via jsDelivr) is the Apache License text, "TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION", Version 2.0.

• https://cdn.jsdelivr.net/gh/langtech-bsc/alia@main/README.md recorded 2026-09-29

  Repository README (via jsDelivr): "Scripts and configuration files used to pre-train the ALIA language model." It also carries an MIT badge, which conflicts with the Apache-2.0 LICENSE file.

• https://github.com/langtech-bsc/alia recorded 2026-06-26

  open training scripts and configuration files for the ALIA models

• https://huggingface.co/api/models/BSC-LT/ALIA-40b recorded 2026-09-29

  Hugging Face model API for BSC-LT/ALIA-40b reports gated as false, carries the tag "license:apache-2.0", and lists weight shards from "model-00001-of-00017.safetensors" through "model-00017-of-00017.safetensors", alongside declared datasets such as "dataset:oscar-corpus/colossal-oscar-1.0".

• https://huggingface.co/api/models/BSC-LT/ALIA-40b/refs recorded 2026-09-29

  Hugging Face refs API for BSC-LT/ALIA-40b lists two tagged earlier revisions of the weights repo, "refs/tags/v0.9" (commit "aa8a4ac7f9e18f3c2ea8ec0cc84e7783cd751ac7") and "refs/tags/v0.5" (commit "6c73c6d901b28aa2fbc17f5267f3b0c1f7b62358"), beside the "refs/heads/main" branch.

• https://huggingface.co/BSC-LT/ALIA-40b recorded 2026-09-29

  Rendered model card states "This model is released under a permissive" Apache 2.0 license, the base model was "pre-trained from scratch on 9.37 trillion tokens", and of the pretraining corpus that it "will not be released or distributed to third parties". Searched the body for intermediate checkpoints: none listed.

• https://huggingface.co/BSC-LT/ALIA-40b/raw/main/README.md recorded 2026-09-29

  Raw model card front matter declares "license: apache-2.0"; the card says "Along with the open weights, all training scripts and configuration files are made publicly available", names the "Colossal OSCAR dataset" and other sources in the data section, and states "The dataset will not be released or distributed to third parties." Searched for intermediate checkpoints: none.

• https://huggingface.co/BSC-LT/ALIA-40b/raw/main/Release%20note%20v0.9.md recorded 2026-09-29

  The v0.9 release note says "the ALIA-40B checkpoint has completed the main pre-training on 8.56 trillion tokens" and that this main pre-training "was unfinished in the previous release", so the tagged v0.5 revision is a partially pretrained checkpoint.

BSC-LT/ALIA-40b's only usage signal is Hugging Face downloads for the single declared checkpoint, and the intended audience is narrow: Spanish and Spain's co-official languages, plus the wider European-language set, rather than a global one. Downloads across the wider ALIA Kit would read higher, but only the base model is scored here.

• https://huggingface.co/api/models/BSC-LT/ALIA-40b recorded 2026-08-13

  475 downloads in the trailing 30 days for BSC-LT/ALIA-40b

• https://huggingface.co/BSC-LT/ALIA-40b recorded 2026-06-26

  ~59 downloads/month, 88 likes (June 2026)

ALIA-40B is a serious 40B multilingual base model, but it is smaller than Apertus-70B and below the 2026 frontier on reasoning and knowledge. It targets Iberian and European-language coverage rather than state-of-the-art capability.

• https://arxiv.org/abs/2502.08489 recorded 2026-08-13

  Salamandra technical report (BSC-LT) describing the shared training methodology and benchmarks

• https://huggingface.co/BSC-LT/ALIA-40b recorded 2026-08-13

  40.43B safetensors parameters, 9.37T training tokens, 35 European languages + 92 programming languages; no benchmark table, with results deferred to a forthcoming technical report