Aleph Alpha releases Kolibri-1: 78B-parameter sparse model (3.46B active), 1M-token context, Apache 2.0

Releases·Tool Release

I squeezed Kolibri-1 78B-A3.5B to 20.9 GiB / 2.30 bpw — 59% lower KL than standard IQ2_XS

A community developer released Sakura-MicroQuality quantizations of Aleph Alpha's Kolibri-1, a ~78B MoE model with ~3.5B active parameters, compressing it to 20.94 GiB at 2.30 bpw in GGUF/llama.cpp format. The quant uses tensor/layer sensitivity to allocate bits where they matter most, achieving ~59% lower KL divergence than a standard IQ2_XS baseline and 90.5% top-token agreement against a Q8_0 reference. A second 365-expert variant prunes 950 rarely-used experts to reach 19.99 GiB at 74.4B parameters with no retraining.

A community developer released Sakura-MicroQuality quantizations of Aleph Alpha's Kolibri-1, a ~78B MoE model with ~3.5B active parameters, compressing it to 20.94 GiB at 2.30 bpw in GGUF/llama.cpp format. The quant uses tensor/layer sensitivity to allocate bits where they matter most, achieving ~59% lower KL divergence than a standard IQ2_XS baseline and 90.5% top-token agreement against a Q8_0 reference. A second 365-expert variant prunes 950 rarely-used experts to reach 19.99 GiB at 74.4B parameters with no retraining.

Social·Community

Aleph Alpha's Kolibri-1 plays Doom with 72 action combinations at 39ms median latency

A demo shows Aleph Alpha's Kolibri-1 model playing Doom, choosing among up to 72 action combinations across four decision groups in a single batch. The demo reports 39ms median and 60ms p95 latency and is playable in a browser.

A demo shows Aleph Alpha's Kolibri-1 model playing Doom, choosing among up to 72 action combinations across four decision groups in a single batch. The demo reports 39ms median and 60ms p95 latency and is playable in a browser.

Social·Community

Aleph Alpha's Kolibri writes working Nim CLI tool on hosted 3code harness

A Reddit post shows Kolibri-1, Aleph Alpha's small German model, writing a working command-line tool in Nim on the 3code coding harness, hosted by Tesseracted. The feat is notable because Nim is a niche language and Kolibri is a small model. All components involved — model, provider, harness, and language — originate from Germany.

A Reddit post shows Kolibri-1, Aleph Alpha's small German model, writing a working command-line tool in Nim on the 3code coding harness, hosted by Tesseracted. The feat is notable because Nim is a niche language and Kolibri is a small model. All components involved — model, provider, harness, and language — originate from Germany.

Releases·Model Release

inclusionAI releases Ling-3.1-flash: 560B-parameter MoE with 25B active and 1M-token context

inclusionAI released the open-source weights for Ling-3.1-flash on Hugging Face, a sparse mixture-of-experts model with 560B total parameters, about 25B active per token, and a 1M-token context window. The jump in scale surprised community members, since its predecessor Ling 3.0 was only 127B total parameters. Practitioners are already discussing local hosting requirements, estimating it needs multiple consumer accelerators to run.

inclusionAI released the open-source weights for Ling-3.1-flash on Hugging Face, a sparse mixture-of-experts model with 560B total parameters, about 25B active per token, and a 1M-token context window. The jump in scale surprised community members, since its predecessor Ling 3.0 was only 127B total parameters. Practitioners are already discussing local hosting requirements, estimating it needs multiple consumer accelerators to run.

Products·building

Kolibri-1 plays Breakout by mapping action probabilities to game moves, no fine-tuning

An experiment using Aleph Alpha's Kolibri-1 model to play Breakout autonomously by interpreting its action probabilities as game moves, without any fine-tuning or generated text. Inference was optimized to about 25ms per decision, demonstrating the model's structured-output ability under constraints. It serves as a capability showcase for researchers exploring small models for real-time decision tasks.

An experiment using Aleph Alpha's Kolibri-1 model to play Breakout autonomously by interpreting its action probabilities as game moves, without any fine-tuning or generated text. Inference was optimized to about 25ms per decision, demonstrating the model's structured-output ability under constraints. It serves as a capability showcase for researchers exploring small models for real-time decision tasks.