Jais 3

# Fix list: Jais 3

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/jais-3, the October 2026 research run, assessed 17 October 2026. Grade D, 51.7 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Jais 3: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Security & auth, 31 out of 100, up to 12.1 more on the total

Why it scored 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), no security contact on the cards, no safety evaluation published, and the technical report is 'coming soon' (4 of 20). 31.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 2. Reliability, 48 out of 100, up to 10.4 more on the total

Why it scored 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 and llama.cpp support is on master only, per the GGUF card (16 of 20). 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 open crash reports. Two template fixes for OpenAI-style message history landed on 14 October, before the launch. Two days is too short to judge (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 and tagged llama.cpp releases don't support the architecture yet (12 of 15). 48.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 3. Schema & documentation, 54 out of 100, up to 7.5 more on the total

Why it scored 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). 54.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 4. Agent ergonomics, 65 out of 100, up to 5.7 more on the total

Why it scored 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). 65.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 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). 60.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 6. Transparency & trust, 50 out of 100, up to 4.4 more on the total

Made of editorial 51, provenance 49.

Why it scored 50: 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 training code, data description and technical report aren't published yet (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). 51.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Domain age: ifm.ai, registered 2022-09-06 (4 years) (7 of 15)
- Status page: not found (0 of 10)
- Changelog: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 62 out of 100, up to 3.3 more on the total

Why it scored 62: Read for an open-weight model released two days before the check. Template changes on 15 October and a citation update on 16 October 2026 (30). One family release, with dated builds before it (FP8 from 29 September, MLX from 1 October, GGUF on 12 and 13 October) and template fixes on 14 and 15 October, but no deprecation notices or successor history to judge. K2 Horizon came on 1 September and Jais 2 in December 2025 (10 of 20). No outside discussions or issues yet on the main repositories, and IFM's GitHub repositories weren't read, so responsiveness can't be seen. The template fixes match the problem IFM's MLX cards describe (7 of 25). Recipes in vLLM's and SGLang's own docs, K2 Horizon support on llama.cpp master and FP8, GGUF and MLX builds from IFM, but no tagged llama.cpp release or native mlx-lm support (11 of 15). Current dependencies (Transformers 5.13 to 5.15, vLLM 0.30), with no CI visible (4 of 10). 62.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: IFM's press release (https://ifm.ai/jais3/press-release/), blog post (https://ifm.ai/blog/jais3/) and sitemap, which answered with a Cloudflare bot check. inceptionai.ai and mbzuai.ac.ae answered with bot or firewall pages. Not gone round
- Launch date. The press release wasn't read. firstReleased is 15 October 2026, the earliest first-hand public evidence. vLLM's recipes repository merged 'Add jais 3 recipes' at 09:14 +04:00 that day, the cards added vLLM recipe links at about 07:44 UTC, and the collections were updated that morning. Gulf News (a lead, not a source) dates the launch to 15 October. The repositories were created from 29 September, and Hugging Face doesn't show when they became public
- unchecked: github.com pages and API, which refused our reader, so llama.cpp pull request 29535 and IFM's GitHub organisation (ifm-ai) weren't read. K2 Horizon support was confirmed in llama.cpp's source on master by a shallow clone
- unchecked: the Weights & Biases training logs linked from the cards, and any IFM evaluation of safety or prompt injection
- Nothing was downloaded or run, so the template errors, parsers and hardware figures are as documented, not observed
- K2 Horizon, IFM's open base family released from 1 September 2026 on the same architecture, has no listing and is a candidate for one. Jais-ASR is a separate speech model and isn't graded here

## Weaknesses

- Every documented vLLM, SGLang, Transformers and MLX command passes `--trust-remote-code`, so the repository's Python runs on the host
- The 375B-A23B reference deployment is 16 H100 GPUs across two nodes, or one 8-GPU H200 node
- The technical report and code repository are marked 'coming soon', and IFM's own blog and press release sit behind a bot check
- No LICENSE file in the repositories, no changelog, no deprecation policy and no security contact or security.txt found
- Released on 15 October 2026, with template fixes on 14 and 15 October, and mlx-lm doesn't yet parse its reasoning or tool calls

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- 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
- Pass `reasoning_effort` (`high`, `medium` or `low`) and `tool_call_format` inside `chat_template_kwargs`. Other values return a template error
- 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
- 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
- Budget at least 32,768 output tokens at `high` effort, which the cards recommend for every request

## What the review panel asked for

- a changelog for template changes
- Publish GPU throughput figures
- Name a hosted provider
- Document structured output
- Correct the 375B effort levels
- upstream the model code
- publish a security contact

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.