Jais 3 vs Prism Inference

Jais 3

Jais 3 runs on the owner's hardware through OpenAI-compatible servers in vLLM and SGLang, with tool calls, three reasoning levels and a 512K context, from a 2.2 GB 0.9B model to a 758 GB 375B one. The weights download without sign-up under Apache-2.0, but every launch command trusts remote code, and no technical report or security contact is published.

Prism Inference

Three wire formats, a public OpenAPI 3.1 file, per-token prices in a keyless catalogue and zero data retention by default on every tier. The service launched on 24 September 2026 with two models, one of them by request, from a two-person company. No rate-limit numbers, SLA document, free tier or deprecation policy was found.

Jais 3

1. Start vLLM or SGLang with --reasoning-parser k2_horizon, and add --tool-call-parser k2_horizon (vLLM also --enable-auto-tool-choice) for tools, or tool calls come back as raw text
2. Pass reasoning_effort (high, medium or low) and tool_call_format inside chat_template_kwargs. Other values return a template error
3. Send each earlier assistant turn back with reasoning_content as a string, even an empty one. A missing or null thinking field makes the template raise an error
4. Send your own system message, or set default_system_prompt to an empty string. Without one the template inserts an identity prompt naming IFM, Inception42 and Cerebras
5. Budget at least 32,768 output tokens at high effort, which the cards recommend for every request

Prism Inference

1. Call GET https://api.prisminference.com/v1/models at start-up, with no key, and use only ids it returns. Expect 403 on gemma-4-31b without organisation access
2. Use base URL https://api.prisminference.com/v1 for OpenAI clients and https://api.prisminference.com with no /v1 for Anthropic clients
3. Read error.retryable before retrying, and wait for Retry-After on 429, which covers both key limits and model capacity
4. Send reasoning_effort: "none" or low when latency matters. Reasoning is on by default and its tokens are billed as output
5. Keep conversation state yourself and send store: false on Responses. previous_response_id, stored responses and hosted tools aren't supported