Aleph Alpha Kolibri Launch: Pricing Impact (October 2026)
Kolibri is a new German-English open-weight AI model. See its license, hardware needs, context window, and what is known about API pricing.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
TL;DR
- Aleph Alpha released Kolibri on October 3: Apache 2.0 weights are downloadable now, but the company has not published a Kolibri per-token API rate.
- It has 78.1B total parameters, about 3.46B active per token, German and English support, and a 262,144-token native context validated to 1,048,576.
- Choose it for self-hosted German/English document work with data-location control; budget GPU capacity and operations, not just tokens.
Cost comparison from today's pricing data
USD per 1M tokens. Input and output rates are charted separately.
Published hosted API rates for alternatives
| Model | Provider | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|
| Mistral Small 4 | mistral | $0.15 | $0.015 | $0.6 |
| Qwen3.5-35B-A3B | novita | $0.25 | n/a | $2.00 |
| GPT-OSS 120B | groq | $0.15 | $0.075 | $0.6 |
Built from pricing.json at publish time.
Aleph Alpha released Kolibri on October 3, 2026. The German company’s launch announcement and model card confirm an Apache 2.0 open-weight model aimed at German and English enterprise work. You can download the weights today. There is no published Kolibri per-token API price in those launch materials, so a direct dollar-per-million-token comparison would be invented.
What changed at launch
| Item | Kolibri Origin | Kolibri 1 |
|---|---|---|
| Release | No public release | Weights available October 3, 2026 |
| Total / active parameters | 30.6B / 3.27B | 78.1B / 3.46B per token |
| Longest trained context | 65,536 tokens | 262,144 tokens; validated to 1,048,576 |
| Reasoning | One mode | None, low, medium, high |
| License | Not publicly released | Apache 2.0 weights and configuration |
These are model specifications, not performance or price guarantees. The roughly 3.5B active parameters make generation relatively compute-efficient, but serving still requires memory for all 78B parameters. Aleph Alpha lists an approximately 78 GB FP8 weight footprint and a minimum of one H200/B200/B300 or two supported 80 GB A100/H100 GPUs. Actual deployment also needs memory for context, concurrency, and runtime overhead.
Kolibri pricing versus hosted alternatives
| Buying option | Published Kolibri token rate | What you pay for |
|---|---|---|
| Download and self-host Kolibri | None; weights are available under Apache 2.0 | GPUs, electricity, storage, engineering, and support |
| Aleph Alpha-hosted Kolibri API | No rate verified in launch materials | Ask Aleph Alpha for availability and a quote before budgeting |
| Existing hosted open-model APIs | See live table below | Provider’s metered input/output tokens |
The live alternative-rate table below is generated from our current pricing dataset. Mistral Small 4, Qwen3.5-35B-A3B, and GPT-OSS 120B are buying benchmarks, not equivalent models or independent Kolibri performance controls. Their published hosted rates do not tell you what Kolibri will cost to self-host. Compare a representative workload’s accepted answers, throughput, latency, GPU utilization, and staff time before claiming savings. See the full Mistral pricing and Groq pricing pages, then use our token calculator to size a hosted baseline. For a managed Qwen control, check Novita’s current Qwen route and verify the exact model and rate before testing.
Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.
Who benefits—and who should wait
German public-sector teams and regulated companies gain a model they can deploy under their own operational control, rather than sending private documents to a third-party inference endpoint. Aleph Alpha says training ran in Germany and Finland and built the model for sovereignty; that is the company’s supply-chain claim, not an automatic certification of any deployment. Kolibri’s bilingual tokenizer and document-grounding training make German-language retrieval and internal knowledge systems sensible first tests.
Small teams without data-center GPUs may find a hosted model simpler to buy and operate. Kolibri’s 1M-token figure is validated extension, not its native training length; the model card recommends at most 262,144 tokens for latency-sensitive and complex tasks. Nor is it an automatic coding-agent winner: Aleph Alpha’s own evaluation reports 27.7 on Terminal-Bench 2.1 against 39.7 for Qwen3.5-35B-A3B. Treat all launch benchmarks as vendor-reported until reproduced on your tasks.
Labs availability
Kolibri is not yet in our maintained Labs roster. The downloadable checkpoint and self-hosted vLLM interface are real, but our benchmark transport has no funded exact route or comparable infrastructure-cost ledger. A defensible 49-task comparison requires a pinned checkpoint and plugin, GPU and precision settings, identical tasks and prompts, raw outputs, latency and failures, token counts, and allocated serving cost. Vendor-reported benchmark scores are not AI Pricing Guru Labs results. See our Labs coverage notes and Aleph Alpha pricing page for the current availability and price boundary.
What to do now
- Download the official weights and model card; confirm Apache 2.0 terms fit your distribution and governance needs.
- Check the Kolibri hardware guidance before reserving GPUs. The Aleph Alpha vLLM plugin exposes an OpenAI-compatible local endpoint, but deployment remains your responsibility.
- Benchmark German/English documents at a 262k-token or smaller context first; expand only after measuring quality and latency.
- If you need managed inference, request a written Kolibri price and service terms. In parallel, compare current open-model rates on our Together AI pricing page and read the local AI versus API cost guide.
Bottom line: Kolibri makes sovereign, self-hosted bilingual AI newly accessible under a permissive license. The pricing story is deployment economics, not a new published API discount; no credible per-token winner can be declared until a hosted Kolibri rate or measured self-hosted cost exists.
Sources: Aleph Alpha launch post, Kolibri product specifications, and official model card. Checked October 3, 2026.