跳转至

Introducing Grok on Amazon Bedrock — Grok 4.3 GA with Mantle Inference Engine

Ch11.257 Introducing Grok on Amazon Bedrock — Grok 4.3 GA with Mantle Inference Engine

📊 Level ⭐⭐ | 5.0KB | entities/grok-43-ga-amazon-bedrock.md

Introducing Grok on Amazon Bedrock — Grok 4.3 GA with Mantle Inference Engine

xAI's Grok 4.3 is now generally available on Amazon Bedrock, marking xAI's entry as a model provider on the platform. Grok 4.3 runs on Mantle, Amazon Bedrock's next-generation inference engine, and uses OpenAI-compatible APIs.

Key Capabilities

Configurable Reasoning Effort

Grok 4.3 supports four effort levels controlled per-request via the reasoning parameter: none (disables reasoning), low (default), medium, and high. Higher effort improves accuracy on multi-step problems at the cost of more output tokens. Classification and extraction can run at none for low latency; contract analysis and case law tasks use high when depth matters.

Model Parameters

  • Context window: 1 million tokens (for long documents and multi-turn sessions)
  • Input: Text and image (PNG/JPEG)
  • Output: Text
  • Defaults: temperature = 0.7, top_p = 0.95, max_completion_tokens = 131072
  • Model ID: xai.grok-4.3

Mantle Inference Engine

Grok 4.3 runs on Mantle, which uses OpenAI-compatible APIs. The endpoint URL is Region-specific (e.g., https://bedrock-mantle.us-west-2.api.aws/openai/v1). Authentication supports both long-term Amazon Bedrock API keys (for exploration) and short-term bearer tokens from IAM credentials (for production).

Tool Calling & Structured Output

  • Supports standard OpenAI tool-calling with JSON Schema parameter definitions
  • tool_choice: "auto" lets the model decide when to call a tool
  • Structured output via json_schema response format with strict mode (additionalProperties: false)
  • One operational note: requests occasionally return 400 from automated content safety checks even on benign input — build a short retry into production calls

Stateful Conversations

The Responses API supports server-side conversation state via store=True and previous_response_id. The service retains each turn's reasoning and feeds it back automatically. Encrypted reasoning is available for stateless use (store=False) with include=["reasoning.encrypted_content"].

Service Tiers & Regional Availability

Tier Description
Standard Pay-per-token, no commitment
Priority Preferential queue processing, higher per-token price
Flex Lower-cost access, not time-sensitive

Grok 4.3 uses in-Region inference only (no Geo or Global cross-Region inference at launch). Example Region: us-west-2.

Comparison with Grok 4.5

Grok 4.3 is the predecessor to Grok 4.5. While Grok 4.3 offers a 1M token context window and runs on Bedrock via Mantle, Grok 4.5 uses a 500K context window with 1.5T parameters on the V9 base. Grok 4.3 is positioned for production workloads that need the full 1M context at lower pricing ($1.25/$2.50 per million input/output tokens vs Grok 4.5's $2/$6).

Practical Guidance

  1. Match reasoning effort to task complexity: Use none for classification/extraction, low for short factual lookups, high for planning steps and multi-step chains where early mistakes derail the task.
  2. Use short-term credentials in production: Long-term API keys are convenient for exploration but should be deleted after testing. Short-term bearer tokens from IAM credentials keep access tied to your identity and expire automatically.
  3. Build retry logic for structured output: The automated content safety check can return 400 on benign input — a short retry loop mitigates this in production.
  4. For image input, validate encoding: Malformed or truncated base64 image payloads return a validation_error rather than a best-guess answer.
  5. Consider service tiers: Standard for pay-per-token, Priority for guaranteed throughput, Flex for cost-sensitive batch workloads.

→ 原文存档