KaziAI-Agent-1.0: Open weights for AI that acts
As our enterprise customers move to KaziAI 2.0, we’re releasing KaziAI-Agent-1.0 under the MIT license. Build business agents on your own infrastructure, with English, Kiswahili and Sheng support.
Most AI models are built to talk. Enterprises need models that act: post the transaction, respect the approval chain, and get the tax right on every invoice. That gap is why we built KaziAI.
Today, we’re opening KaziAI 1.0
Our enterprise customers are now moving to KaziAI 2.0. So today we’re releasing the weights of KaziAI-Agent-1.0 under the MIT license.
The release gives developers a business-specialized model to connect to their own tools and infrastructure. In the KaziAI agent system, the model plans structured tool calls while the surrounding runtime handles exact finance math, session state and access enforcement.
What KaziAI 1.0 does
With connected business tools and the KaziAI runtime, the agent system is designed to:
- Execute ERP, CRM, POS and clinical-admin workflows through structured tool calls.
- Hold transaction state and entity identifiers across 20 to 50 turn sessions, including interrupted ones, using the runtime’s session ledger.
- Enforce role-based access through the runtime: payments of $1,000 or KES 100,000 and above route to supervisor approval, and destructive actions such as customer deletion are refused.
- Isolate tenant data at the parameter level with runtime authorization checks.
- Compute East African compliance deterministically through the finance engine: 16% VAT, 2% withholding VAT, 5-band PAYE and M-Pesa reconciliation.
- Stop and request clarification on ambiguous instructions instead of assuming.
- Work in English, Kiswahili and Sheng, so frontline teams can use it the way they speak.
- Deploy on your infrastructure: the Q8_0 model uses approximately 2.1 GB of active RAM with a 4,096-token context; vLLM and multi-tenant LoRA serving are also supported.
Under the hood
- DeepSeek-R1-Distill-Qwen-1.5B base with a merged rank-128 LoRA, trained on 8,000 multi-turn enterprise trajectories.
- Trained to reason through permissions, missing parameters and business constraints before emitting tool calls.
- A neuro-symbolic runtime around the model handles exact finance math, session state and RBAC interception.
Reported evaluation results
These are self-reported figures from the release model card. Business-workflow and combined security scores include the agent runtime.
- 94.11% BFCL AST match; Qwen-2.5-7B-Instruct scored 88.42% in our reported comparison runs.
- 98.44% task completion on our Track B business simulation of 2,500 tasks.
- 94.20% on 20 to 50 turn interrupted enterprise workflows, as reported in the model card.
- Across 1,500 adversarial attempts, the model card reports 96.8% standalone refusal and 99.73% overall containment with the runtime policy layer.
One repository. Choose how you run it.
The release includes merged safetensors, Q8_0 and BF16 GGUF, an Ollama Modelfile and the raw LoRA adapter. Load the merged checkpoint with Hugging Face Transformers, run locally with Ollama or llama.cpp, or serve with vLLM.
Build with KaziAI
If you’re building agents for African businesses, start here. For what our enterprise customers now run on, talk to us about KaziAI 2.0.