The 2601 build offers a 163,840-token window on a dense 40B model that BSC aligned with preference training. It suits teams that want a self-hosted, Apache-licensed assistant for Iberian languages, and it is a poor fit for agents, tool calling or maths.
What changed
ALIA-40b-instruct-2601 is a 40.4B-parameter open-weights assistant model from the Barcelona Supercomputing Center, published in January 2026. BSC's card reports 45% on MMLU in a 0-shot run and a 13.3% average red-team attack success rate, with post-training focused on Spanish, Catalan, Basque, Galician and English.
Provider: Barcelona Supercomputing Center · Family: ALIA
More about Barcelona Supercomputing Center on HokAI
Context window: 163,840 tokens
Input modalities: text · Output: text
About ALIA-40b-instruct-2601
ALIA-40b-instruct-2601 is the instruction-tuned release in the ALIA family from the Barcelona Supercomputing Center (BSC), Spain's national supercomputing centre. It is a dense, decoder-only Transformer with 40,433,885,184 parameters, 48 layers, 64 attention heads with grouped-query attention and a 256,000-token vocabulary, published under the Apache 2.0 licence. It is the instruct variant of ALIA-40b, a base model BSC pre-trained from scratch, and per BSC's model card it improves on the earlier ALIA-40b-instruct-2512 in instruction following, input robustness, safety alignment and long-context behaviour.
The base model saw 9.37 trillion tokens of curated text in 35 European languages plus code, then 424 billion more high-quality tokens and 39 billion tokens from a similar mixture, for 9.83 trillion in total. Post-training has three stages. Long-context adaptation first stretched the window (2 billion tokens at 4k to 32k, then 36.8 billion at 32k to 160k). Supervised fine-tuning followed on 807,750 conversations, and preference alignment used Direct Preference Optimization on 368,475 pairs, about 82% aimed at helpfulness and 18% at safety. BSC states that much of the instruction and preference data is synthetic and was generated or judged with DeepSeek-V3-0324, a model from DeepSeek. Training ran on BSC's MareNostrum 5, whose accelerated partition has four NVIDIA Hopper GPUs per node; each post-training stage used 16 nodes and 64 GPUs.
The context window is 163,840 tokens and both input and output are text. BSC lists text generation, translation, summarization and question answering as the target tasks. Post-training concentrated on Spanish, Catalan, Basque, Galician and English, and BSC says performance in other languages is not guaranteed. The card also states plainly what the model does not do: it has no native function calling, no agent or autonomous workflow capability, no guarantee of reliable multi-step maths, and no specialisation in code.
BSC reports 0-shot results from EleutherAI's lm-eval suite. In English the card gives 45% on MMLU, 77% on Belebele, 73% on ARC-Easy and 9% on MGSM. In Spanish it gives 72% on Belebele and 41% on MMMLU. These figures are vendor-reported and not independently reproduced, BSC warns of about 1.5% variance depending on library version and tensor parallelism, and they are not comparable with the five-shot MMLU numbers quoted for other models on HokAI.
On safety, BSC ran a red-team test with adversarial prompts written by its own annotators in five languages, using DeepSeek-V3-0324 as the judge, and reports an average attack success rate of 13.3%. It describes alignment as substantially improved but incomplete and tells deployers to add their own filters, human oversight and evaluation before any critical use.
The weights are on Hugging Face as BF16 safetensors. BSC recommends vLLM for the full model and llama.cpp for the quantized GGUF build, and it discourages Ollama because of compatibility problems that can degrade output. It advises temperatures of 0 to 0.2 with no repetition penalty. BF16 weights alone come to roughly 81 GB (40.4 billion parameters at 2 bytes each) before the KV cache.
Choose it for self-hosted assistants, translation and summarization in Iberian languages where an Apache licence and a public-sector publisher matter. Choose Gemma 4 31B from Google DeepMind or Qwen3.8 27B from Alibaba Cloud when you need tool calling and a 262,144-token window, and Olmo 3.1 32B Think when published training data is the requirement. The ALIA work is funded by the Spanish Ministry for Digital Transformation through the NextGenerationEU project Modelos del Lenguaje and supported by the Government of Catalonia through the Aina Project.
Pricing
BSC runs no hosted API, so there is no per-token price. Downloading the weights costs nothing; the bill is whatever hardware you run them on. Hosting on [Hugging Face](/hub/tools/hugging-face) Inference Endpoints is one route, billed by Hugging Face.
| Tier | Rate |
|---|---|
| Self-hosted (Apache 2.0) | Free weights from Hugging Face; cost is hardware and electricity only. |
| Hugging Face Inference Endpoints | Dedicated endpoint you deploy from the model page; billed by Hugging Face, price depends on the instance. |
Key Features
- Two-step long-context adaptation: BSC extended the base model from 4k to 32k tokens and then to 160k, in two documented steps, which BSC credits for the long-context gains over the 2512 build.
- Iberian-language instruction data: The supervised mixture holds 103,815 Catalan, 72,011 Basque and 71,123 Galician conversations beside 161,390 Spanish and 342,982 English ones.
- Open training recipe: Weights, training scripts and configuration files are public, and the card tabulates the licence of every supervised dataset.
- Preference alignment with safety data: The alignment set has 368,475 preference pairs, and the safety prompts cover nine harm categories in five languages.
Pros
- No usage fees and no vendor lock-in: Apache-licensed weights run on your own hardware and your prompts stay in-house.
- Instruction data is itemised by language and licence, which helps an audit of how the model was built.
- A named public-sector publisher and funding source (NextGenerationEU, the Aina Project) can simplify provenance questions in procurement.
Cons
- No native function calling or agent loops, according to BSC's own limitations list.
- Maths and multi-step reasoning are weak: the card shows single-digit MGSM scores.
- Needs multi-GPU hardware or quantization, since the BF16 weights alone are around 81 GB.
Benchmarks
- MMLU: 45% vendor-reported · 12 Oct 2026 — General-knowledge exam across 57 subjects, % correct.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much do you pay for ALIA-40b-instruct-2601?
Nothing for the weights: BSC publishes them under Apache 2.0 and runs no per-token API. Your cost is GPU time, and the full-precision model needs multi-GPU memory. Hugging Face Inference Endpoints can host it, with the price set by Hugging Face and the instance you choose.
Is ALIA-40b-instruct-2601 free to use?
Yes. The licence allows commercial use, and BSC's card describes the family as meant for research and commercial work. The limit is risk, not cost: BSC says deployers must test the model and comply with applicable AI regulation, and it warns against critical applications without extensive testing.
What should you use instead of ALIA-40b-instruct-2601?
For tool calling, pick Gemma 4 31B or Qwen3.8 27B. For fully published training data, pick Olmo 3.1 32B Think. For a smaller Iberian-language model on modest hardware, BSC's own Salamandra 7B Instruct has an 8,192-token context.
Is ALIA-40b-instruct-2601 better than Gemma 4 31B?
For general work, no. Gemma 4 31B accepts images, supports tool use and has a thinking mode, while BSC's card lists no function calling or agent support. ALIA wins only when Catalan, Basque or Galician coverage and a public Iberian publisher matter more than those features, so test both on your own text.
How long does it take to get going with ALIA-40b-instruct-2601?
On a machine with enough GPU memory, the Transformers snippet in BSC's card runs once the download finishes, so the wait is mostly the download and load. The Deploy button on Hugging Face builds a vLLM endpoint in a few clicks, which BSC recommends as a first deployment. Set temperature near 0.1 and leave repetition penalty off.
Top Alternatives
- Gemma 4 31B: Pick Gemma 4 31B for image input, tool use and a thinking mode; pick ALIA for Catalan, Basque and Galician post-training from a public publisher.
- Qwen3.8 27B: Pick Qwen3.8 27B for a longer window and tool calling; pick ALIA when Iberian-language coverage and documented instruction data matter more.
- Olmo 3.1 32B Think: Pick Olmo to audit published training data; pick ALIA for Spanish and co-official Spanish languages with an instruct-first design.