Amazon Bedrock vs Jais 3

Amazon Bedrock 98

Model reading. IAM with SigV4, roles and STS session credentials, policies down to a model or inference-profile ARN, short-term API keys that expire within 12 hours, long-term keys with expiry and reset, and condition keys that can deny bearer tokens altogether (30). The security page says customer data isn't shared with model providers or used to improve the base models (20). The abuse detection page says Bedrock stores no inputs or outputs by default, and a per-Region none mode blocks models that require retention, but Claude Fable 5 and 5.1 keep all traffic up to 30 days and GPT-6 and GPT-5.x keep flagged traffic up to 30 days (13 of 15, a judgement call; Anthropic's own API got 12 for the same Fable exception without an account-level block). CloudTrail logs API calls, and model invocation logging can send prompts and responses to S3 or CloudWatch Logs (15). Bedrock is in ISO, SOC and CSA STAR Level 2 scope, HIPAA eligible and FedRAMP High in GovCloud (US-West) per its security page, with a vulnerability disclosure programme on HackerOne, and the aws.amazon.com security.txt, expired on 6 October, was renewed with Expires 2027-10-16 when read on 6 November (20). Up from 96 on 6 October for that reason.

Jais 3 31

Read as software the owner runs. No account or key, ungated downloads, and server authentication is the owner's to add (12 of 30). A model that calls whatever tools the operator wires in. The cards rule high-stakes decisions without human oversight out of scope, but nothing in the model or template enforces approval (6 of 20). The template renders tool results in their own tool role, but string results go in as raw text (only non-string content is serialised with tojson), no escaping of special-token sequences in caller or tool text was found, and no injection guidance or results are published (4 of 15). No call log of its own. Visibility depends on the serving stack (5 of 15). Every launch command passes --trust-remote-code, so modeling_k2_horizon.py from the repository runs on the host. We read it and found only torch and transformers imports. No security.txt on ifm.ai (404 on 17 October, not re-read), no security contact on the cards and no safety evaluation found. Since 25 October the cards link a technical report on ifm.ai, which answered a Cloudflare bot check on 4 November, so what it covers is unread (4 of 20). 12 + 6 + 4 + 5 + 4 = 31.

Amazon Bedrock 94

Model reading. No OpenAPI, but the bedrock-runtime and bedrock Smithy models are public in aws/api-models-aws with types, required members and enums, counted as for S3 (25). llms.txt for the user guide with about 2,000 lines, each page with a Markdown twin (10). Model cards state context, output limits, endpoints and supported functions per model, and the guide now recommends bedrock-runtime for new work, but support for each API, endpoint and feature changes model by model (16 of 20). Structured outputs compile JSON Schema to a grammar, with minimum, maximum and string lengths supported, though some inference profiles don't support it, such as Haiku 4.5 through the India profile (13 of 15). Every operation lists typed errors with HTTP codes and links to a troubleshooting page with causes and fixes, and examples run through the guide (15). API version 2023-09-30 and a dated document history with an RSS feed, newest entry 30 October 2026, with structured outputs for Claude on both endpoints added on 18 October (15).

Jais 3 54

Read for a model the owner serves. No specification of its own. The contract is the Jinja chat template, which validates tool JSON Schemas (required fields must exist in properties) and rejects unknown options, plus the OpenAI-compatible servers in vLLM and SGLang with the k2_horizon parsers (10 of 25). The cards are Markdown on Hugging Face, and SGLang's docs have an llms.txt, but no llms.txt or Markdown docs from IFM were found and ifm.ai couldn't be read (5 of 10). Each card states intended use, four out-of-scope uses, limitations and per-size serving advice (15 of 20). reasoning_effort and tool_call_format are enums enforced with errors, and tools take JSON Schema. The 375B card describes low as no reasoning and omits medium, which the template accepts, and structured output isn't documented (10 of 15). Python examples for a reasoning call and a tool loop, commands for five runtimes, and named template errors, but no error reference for a served model (11 of 15). No changelog or version scheme. Training-stage branches and public commit history (3 of 15). 10 + 5 + 15 + 10 + 11 + 3 = 54.

Amazon Bedrock 80

AWS Health Dashboard with per-service, per-Region feeds and a public history file (20). The history lists Amazon Bedrock as disrupted in ME-CENTRAL-1 (UAE) and ME-SOUTH-1 (Bahrain) since March 2026, where AWS said on 15 September 2026 (still inside the 90-day window on 6 November) that it can't restore resources held only there. No other Bedrock event appears in the history in the 90 days to 6 November; an EU-SOUTH-2 packet-loss event on 4 October named Bedrock AgentCore, not Bedrock. This counts as one major event (10). StatusGator's third-party mirror shows 'Increased Error Rates' warnings on 24, 26 and 27 August and 10 September that we couldn't find in AWS's own history, so we didn't score them. Per-model token quotas with default values in the AWS General Reference, such as 30,000,000 tokens a minute for Opus 5.5 global inference (15). The error guide gives exponential backoff with jitter for 429, 503 and 529 and says to honour Retry-After, and the SDKs retry ModelNotReadyException up to five times (15). SLA of 99.9 per cent a month per Region for the model APIs, last updated 4 October 2023 (10). Converse and the current Claude, GPT-6 and Nova 2 Lite models are GA (10).

Jais 3 48

Read with the local-software lines, since the owner serves the weights. Ungated weights on Hugging Face with stated runtimes (vLLM 0.30.0 or later, SGLang 0.5.20 or later, Transformers validated at 5.15.0, mlx-lm with shipped model code), but every launch command trusts remote code. The GGUF card still says to build llama.cpp from master or wait for a release, though llama.cpp's own tags carry K2 Horizon support from b11454 (6 October 2026) (16 of 20, held). No CI or test suite of the makers' was found, and the code repository is 'coming soon'. SGLang's cookbook reports a validation pass of the tool-call parser on all six checkpoints, and IFM's MLX cards describe conversion checks against the PyTorch code (3 of 25). No crash reports. On 4 November the six main repositories held seven discussions, six of them closed citation updates and one open MoVA-36B-A4B thread from 26 October with no reply yet. Two template fixes for OpenAI-style message history landed on 14 October, before the launch (15 of 25). No changelog or version tags, only public commit messages (2 of 15). Released as named final checkpoints with no preview label, though mlx-lm v0.32.2 still has no K2 Horizon model code of its own (12 of 15). 16 + 3 + 15 + 2 + 12 = 48.

Amazon Bedrock 30

No x402, MPP or L402 (0). Per-token prices per model, tier and Region on the public pricing page, filled from AWS's public price feed (20). New accounts get up to $200 in Free Tier credits and the Free Tier page lists Bedrock among eligible services, and the Free Tier FAQ says most new customers don't need a payment method, but AWS may still ask for one, so half, as for S3 (10). A person signs up for AWS in a browser and sets up IAM before any key exists (0).

Jais 3 60

Self-hosted rule. Apache-2.0 weights with nothing to buy and no hosted option from the makers (20). Free to run with no card (20). Downloads are ungated and anonymous, so an agent can fetch the weights with no sign-up (20). No payment protocol (0). 20 + 20 + 20 + 0 = 60.

Amazon Bedrock 85

Closed service under the AWS Customer Agreement and Service Terms, with the SDKs and Smithy models under Apache-2.0 (18, as for S3). The data retention, abuse detection and security pages agree that Bedrock stores nothing by default, name the models that retain traffic and for how long, and say retained data isn't shared with providers. The retention page's default mode says AWS 'may retain the data for safety and abuse-prevention purposes' without a period, and a legacy provider_data_share mode is kept though sharing isn't supported (25 of 30). A written lifecycle policy with Legacy and EOL dates per model, but the legacy table's 34-day Legacy periods for five Claude models in some Regions contradict its own 6-month floor (15 of 20, down from 18). AWS's sub-processor list, per the 4 October S3 check, and per-Region, geographic and global inference options, with global routing able to process data in any commercial Region (18 of 20).

Jais 3 55

Apache-2.0 in every card's metadata, with staged training checkpoints as branches and training logs linked on Weights & Biases (not opened). There is no LICENSE file and the code repository is still 'coming soon'. A technical report is linked from 25 October but sits behind a bot check, so its data description is unread (24 of 30). Inputs stay on the owner's hardware because nothing is hosted, but no statement says so and no privacy document governs the weights (15 of 30). No deprecation policy, dated notices or version scheme (0 of 20). Safetensors weights carry no code, and the remote modelling and configuration code we read imports only torch, transformers and huggingface_hub, with no network calls. No telemetry statement (12 of 20). 24 + 15 + 0 + 12 = 51.

Amazon Bedrock 92

Model reading of the checklist (tool use, structured output, caching, context, batch, SDKs, errors). Converse takes tools with toolChoice auto, any or a named tool and strict: true, but forced choice depends on the model behind it (17 of 20). Structured outputs on supported models only (13 of 15). Explicit and implicit prompt caching, cache reads at $0.20 per million on Opus 5.5, 0.05x input, with 5-minute and 1-hour TTLs (15). 1M-token context on Opus 5.5, Sonnet 5.5 and Nova 2 Lite, a judgement call since Bedrock has no default model (15). Batch at half price on supported models, but the pricing page shows none for Opus 5.5 or Sonnet 5.5, while Flex is half price on Nova (7 of 10). AWS SDKs in every major language plus token generators for Python, JavaScript and Java (10). Typed errors with recovery advice per code (15).

Jais 3 65

Model reading of the checklist (tool use, structured output, caching, context, throughput, defaults and SDKs, errors). OpenAI-format tools come back as tool_calls in three selectable wire formats, with a multi-turn loop example. Forced tool_choice and parallel calls aren't documented (15 of 20). JSON mode and schemas only through each runtime's guided decoding, not documented by the makers (5 of 15). Prefix caching is the runtime's, and SGLang's page covers HiCache (7 of 10). 524,288 tokens on five of six models, 131,072 on the 0.9B (14 of 15). Uno adapters claim 2.15 times faster decoding for 0.9B to 32B but need IFM's own engine. SGLang lists n-gram speculative decoding and PD disaggregation. No batch mode of its own (7 of 10). Any OpenAI SDK works against a local server, and the 0.9B to 7B fit one GPU. But the cards ask for high effort and 32,768 output tokens on every request, the template inserts an identity system prompt when none is sent, and earlier assistant turns must carry a string thinking field (8 of 15). The template raises named errors for bad options, string-encoded tool arguments and malformed schemas. Server errors are the runtime's (9 of 15). 15 + 5 + 7 + 14 + 7 + 8 + 9 = 65.