Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Aleph Alpha puts Kolibri's weights on Hugging Face, and llama.cpp learns to answer without writing

Aleph Alpha released Kolibri, a 78B-parameter open-weight model with 3B active parameters, on 3 October, the same day llama.cpp added support for decision models that return probabilities instead of text.

AI & modelsAnalysisRachel NwosuPublished: 3 October 20264 min readSources 9
Aleph Alpha puts Kolibri's weights on Hugging Face, and llama.cpp learns to answer without writing

Aleph Alpha published Kolibri on 3 October. The model is an English-German Mixture-of-Experts Transformer with 78B total parameters and 3B active, and the company put the full weights on Hugging Face under Apache 2.0. It takes context windows of up to 1M tokens.

The timing matters for buyers in regulated sectors. Kolibri is pitched at public administration, industrials and aerospace, and Aleph Alpha says customers can run it on-premise so internal data never reaches a third-party inference service. No price or customer count was disclosed.

The company frames the release around supply-chain integrity: it says it can account for every step from data ingestion through pre- and post-training to final evaluations. Kolibri follows an earlier 30B model, Kolibri Origin, built on the same pipeline. Aleph Alpha claims Kolibri matches models with up to four times its active parameter count, naming Nemotron 3 Super as an example, and says it sits on the Pareto frontier for quality versus serving cost in both English and German. Those are vendor benchmarks run on the company's own internal evaluations, not independent ones.

Weights are getting cheaper to ship, and smaller to run

The same weekend brought a second kind of open release. It is the more interesting one. The llama.cpp project said on 2 October that its server now supports decision models through a /v1/systemone endpoint, per a Hugging Face community post by Xuan-Son Nguyen and Victor Mustar. You send a state, text, JSON or a screenshot, plus typed questions, and the model returns a probability for each option after a single forward pass. No tokens are generated. Existing clients that speak TypeSafe's Jev format only need a new base URL.

The model table in that post is the part worth reading closely. Laya, a 421M-parameter English model under Apache 2.0, answers one question in a median 5 ms on an NVIDIA RTX PRO 6000. Clef is 27B, takes images, and is Apache 2.0. OpenJev is 27B and multilingual but ships under CC BY-NC 4.0, which rules out commercial use. Kev-4B answers in 12 ms, lev in 36 ms. A decision model that routes a support ticket or checks an agent's step does not need to write prose, and the cost difference follows from that.

Kitsuno, which runs Jev in production, published its own numbers on 3 October and they cut both ways. Since 23 September Jev answered 35,812 logged calls for 76 million input tokens at a total cost of USD 3.68. On a test set built the same day, untrained Laya answered 44 percent of questions correctly against Jev's 94. Kitsuno says it did not train Laya on Jev's output because TypeSafe's customer agreement, section 2.3(b) as updated on 23 September, prohibits training a similar or competing product on the service's output. It calls that its reading of the contract, not legal advice.

Two smaller open releases landed alongside. ZeroLeaks published Shield Small, a 118M-parameter classifier for prompt injection and jailbreak detection, on 2 October, with int8 ONNX weights of 118,445,560 bytes that run on a CPU. Its model card reports 84.3 percent mean category-balanced accuracy and 71.8 percent attack recall when paired with Shield rules v2, and notes the comparison covered 35 public datasets that informed development rather than an unseen test. The same caveat applies to the false-positive rate of 7.2 percent.

Not everyone agrees open weights make things safer

Anthropic used its own report to argue the opposite direction. Tom's Hardware reported on 3 October that Anthropic claims Zhipu AI's GLM-5.3 can generate malicious content with weak safeguards. In Anthropic's Exploitbench runs, GLM-5.3 produced end-to-end Chrome exploits 50 times in 410 attempts, against 56 for Anthropic's unreleased Mythos model. On an internal benchmark for full control-flow hijacks it scored 4 percent against Mythos at 6 percent, while Kimi K3 and DeepSeek V4.1 Flash both scored 0 percent. Anthropic says a deceptive prompt bypassed the guardrails 64 percent of the time and prefilling thinking tokens worked 92 percent of the time. These are Anthropic's numbers about a competitor's model, and Zhipu AI has not responded in the material reviewed here.

The policy fight is not settled either. WIRED reported on 3 October that Nathan Lambert and Tom Zick launched Trillium Labs, a nonprofit that will publish details of experiments in areas including recursive self-improvement and agents.

Comments 0

Sources

9
  1. 01Kolibri Has Landed: A Sovereign Open-Weight ModelEN
  2. 02New in llama.cpp: Decision ModelsEN
  3. 03Laya in shadow: training an open decision model to replace JevEN
  4. 04Shield: A 118M model for detecting prompt injections and jailbreaksEN
  5. 05Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
  6. 06These AI Experts Want to Do High-Stakes Research Out in the OpenEN
  7. 07Anthropic's super bug-hunting model Mythos is hardcore good at mathEN
  8. 08As Part of Ongoing Investigation, Attorney General Bonta Serves Investigative Subpoena on OpenAIEN
  9. 09OpenAI safety leader David Robinson resignsEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.