Models

Tase serves purpose-built models rather than one generic model for everything. Each model in the family is trained by Tase Industries for a specific role in the product, so you pick the model that matches the job instead of paying for capability you do not use.

Available models

• tase-0.1-8b — the default model. Trained by Tase Industries and specialized for tool calling inside the Tase app: picking the right tool, producing correct arguments, and responding with low latency in both Turkish and English. Supports chat completions, function calling, and streaming.

Choosing a model

• Tool-driven actions: creating a task, logging an expense, checking a schedule — use tase-0.1-8b. This is what it was trained for.
• Everyday conversation: quick questions and short answers in Turkish or English — tase-0.1-8b handles these with the lowest latency in the family.
• Complex, long-form reasoning: send the request to tase-0.1-8b anyway — routing takes care of escalation automatically. You do not need to select a different model yourself.

Routing

Some requests need deeper reasoning than a compact model should attempt. When Tase detects a complex, long-horizon request, it can route that request to a larger model transparently. The API surface stays the same: you address tase-0.1-8b, and routing decides whether escalation is needed. Responses arrive on the same connection in the same format, so your integration does not change.

Listing models

Call GET /v1/models to see the models available to your account. Availability can differ by plan.

curl https://api.tase.app/v1/models \
  -H "Authorization: Bearer $TASE_API_KEY"

{
  "object": "list",
  "data": [
    {
      "id": "tase-0.1-8b",
      "object": "model",
      "owned_by": "tase-industries",
      "capabilities": ["chat.completions", "function_calling", "streaming"]
    }
  ]
}

Industry benchmark

On the public Berkeley Function Calling Leaderboard (BFCL v4), Tase 8B scores 87.58% on non-live AST function calling — second overall across every evaluated model, behind only Claude Opus 4.5, and the top of the open 8B class. Tase has not been submitted to BFCL as a separate entry; the rows below reflect the open 8B model Tase is built on.

Across all five BFCL metrics

Non-live AST is one slice. BFCL v4 also grades live functions, irrelevance detection (correctly declining to call a tool), multi-turn accuracy, and an overall aggregate. Tase 8B stays at or above the frontier-model median on the two that matter most for a personal assistant — picking the right tool on a single turn, and knowing when not to call one.

8B open class head-to-head

Against the only other open 8B model on the leaderboard — Llama 3.1 8B from Meta — Tase 8B leads on every BFCL v4 axis. The widest gap is irrelevance detection (79.07% vs. 42.70%): Tase is far less likely to call a tool when the user just wanted an answer.

Tase 8B at a glance