Aleph Alpha
Aleph Alpha's Kolibri lets users dial reasoning effort from none to high
Aleph Alpha's recently released Kolibri model lets users choose how long it reasons, with four settings: none, low, medium or high. The company says its training data…
Aleph Alpha distills its parsing harness into Parsy, a single-pass model
Aleph Alpha has trained a specialized model called Parsy to replace its own multi-pass agentic parsing harness. According to the company, Parsy scores 87.9 on ParseBench…
Aleph Alpha: Machine-Translated German AI Benchmarks Often Measure the Wrong Thing
Aleph Alpha says many German versions of popular LLM benchmarks do not measure what the English originals were built to test. After reviewing thirty German-translated…
One Space Shifts Answer Probabilities by Orders of Magnitude, Aleph Alpha Finds
Where a single space sits in a multiple-choice prompt can change the probability a model assigns to the correct answer by several orders of magnitude, according to new…
ResearchAleph Alpha details two-scale expert balancing behind its Kolibri 1 MoE model
Aleph Alpha has published a research post on the load-balancing recipe used to pre-train Kolibri 1, a sparse Mixture-of-Experts model. The method combines Exact Quantile…
Aleph Alpha: Bits-per-Byte Metric Penalizes Short Answers, PriorBPB Fixes It
Aleph Alpha researchers say the widely used bits-per-byte (BPB) metric for evaluating pre-trained language models is biased by the length of the correct answer, and they…
Aleph Alpha says its Kolibri model beats larger rivals on agentic RAG tasks
Aleph Alpha says its open-weight model Kolibri achieves the best average score across seven agentic RAG benchmarks among the leading open-weight models the company…
ResearchAleph Alpha details how architecture choices drive LLM cost, with Kolibri as reference
Aleph Alpha has published a research blog post that works out how core architecture choices in autoregressive language models determine training and serving cost. The…
ResearchAleph Alpha: Leaked Benchmark Data Inflated HumanEval Scores by 45 Points
Aleph Alpha says it removed 1.5 million documents quoting benchmark items from the mid-training pool of its Kolibri model, and that putting them back in a proxy run…
Aleph Alpha releases Kolibri, a German-focused model for public sector and industry
Aleph Alpha has released Kolibri, a specialized large language model for mission-critical applications in public administration and industry. According to the company…
ModelsAleph Alpha releases Kolibri, an open-weight 78B model for regulated sectors
Aleph Alpha has released Kolibri, an English-German Mixture-of-Experts Transformer with 78B total parameters and 3B active, under the Apache 2.0 license. According to…
ResearchAleph Alpha Scales 30B-A3B MoE Pre-Training to 512 B200 GPUs at 35.3% MFU
Aleph Alpha has published a research post explaining how it scaled pre-training of a 30B-A3B mixture-of-experts model from 16 to 512 NVIDIA B200 GPUs. According to the…