Alpha
Contact
Implement · Jais 3

Can you run it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 57.6/100

This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Install & run

The weights are ungated and need no login (ev-hub). Load them with Transformers 5.15.0 (the version the cards validate) and trust_remote_code=True, or serve them with vLLM or SGLang.

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "inception42/Jais-3-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", dtype="bfloat16", trust_remote_code=True
)

The card example drops token_type_ids from the tokenizer output before generate (ev-card-32b). Pass revision="<commit-sha>" or a stage branch name to pin a version. Read configuration_k2_horizon.py and modeling_k2_horizon.py before first use, because they run on your machine under trust_remote_code=True.

Hardware & VRAM requirements

The cards give tested SGLang layouts on H100 GPUs, with BF16 and FlashAttention-3 (ev-card-32b).

ModelTested layoutBF16 weights (from safetensors parameter counts)
0.9B, 3.7B, 7B1 GPU (TP=1)about 2 GB, 10 GB, 18 GB
32B2 GPUs (TP=2)about 70 GB
MoVA-36B-A4B2 GPUs (TP=2, EP=2)about 75 GB
375B-A23B16 H100 across two nodes (TP=16, EP=16), or one 8xH200 node (TP=8, EP=8)about 758 GB

The weight sizes are OneHill arithmetic from the hub parameter counts (ev-hub). No figure is published for the KV cache at 512K context, which adds a large amount of memory. FP8, GGUF and MLX builds reduce weight memory (ev-quants).

Serving stacks

The cards document three paths (ev-card-32b).

vLLM 0.30.0 or later:

vllm serve inception42/Jais-3-32B --trust-remote-code --model-impl vllm --dtype bfloat16 \
  --tensor-parallel-size 2 --reasoning-parser k2_horizon \
  --enable-auto-tool-choice --tool-call-parser k2_horizon

SGLang 0.5.20 or later with --reasoning-parser k2_horizon and --tool-call-parser k2_horizon. Transformers 5.15.0 with PyTorch 2.13.0 and Safetensors 0.8.0.

GGUF builds target llama.cpp and Ollama according to the cards. MLX builds target Apple silicon. OneHill has not run any of these stacks.

Safe-deployment controls & Deployment Ceiling

Ceiling: T1 (internal assistant, human reads every output), conditional. Every safety result is the publisher's own, no text guard model is released, and no result covers tool, terminal or file actions. Customer-facing use (T2) needs your own evidence first.

To reach T2, complete these steps.

  • Run your own jailbreak and misuse evaluation in the dialects you serve, on the exact revision you deploy.
  • Add input and output guardrails and a guard classifier of your own.
  • Pin the revision, verify the LFS sha256 values and read the two custom Python files.
  • Sandbox every tool the model can call. Agent behaviour is the model's main design goal, and its safety has no published test.
  • Record that you deploy an Apache-2.0 model whose card tells users to add safeguards before deployment (ev-cards-all).

Available quantizations

The IFM organisation publishes official builds (ev-quants).

FormatSizesNotes
FP83.7B, 7B, 32B, MoVA-36B-A4B, 375B-A23Bno 0.9B build
MLX 4-bit, 8-bit, bf160.9B, 3.7B, 7B, 32B, MoVA-36B-A4BApple silicon
GGUF0.9B, 3.7B, 7B, 32B, MoVA-36B-A4Bthe 32B repository holds BF16, Q4_K_M, Q5_0, Q5_K_M, Q6_K and Q8_0; none for the 375B

GGUF files appeared on 2026-10-12 and 2026-10-13, two weeks after the weights. The Uno LoRA adapters speed up decoding, but they carry no licence tag (ev-uno). Community quantisations were not assessed.

Fine-tuning & adaptation

Every size has branches for the CPT, Cooldown, FTPO and SFT checkpoints, so you can continue from any stage (ev-branches). The xLLM repository covers pretraining and SFT, RL360 covers RLVR, and the data sample covers four Arabic subsets (ev-code, ev-data-sample).

No Jais-3 run configuration was found in xLLM or RL360, whose READMEs describe K2 Horizon. The report lists configurations and W&B logs as released, but they were not opened. Check them before planning to reproduce a stage.

Keep the model's chat template, include reasoning-format samples, and mix in your own safety data. Further training on new data can erode safety behaviour, as the report itself notes for continued training of post-trained models.

API / OpenAI-compatible integration

vLLM and SGLang expose OpenAI-compatible /v1/chat/completions endpoints (ev-card-32b). Set reasoning depth per request:

extra_body={"top_k": 20, "chat_template_kwargs": {"reasoning_effort": "high"}}

Reasoning is returned in reasoning_content and the answer in content. For tools, start the server with --enable-auto-tool-choice --tool-call-parser k2_horizon and pass tool_call_format (xml, json or xml_typed) in chat_template_kwargs. The card advises reasoning_effort="high", at least 32,768 output tokens, temperature 0.7 for Arabic thinking and 1.0 for English thinking.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Weak

Safety scores 2: every result is the publisher's own, there is no guard model, and no test covers agent actions. Operational scores 3 (ev-report-safety, ev-card-32b, ev-hub).

How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionThe cards give validated vLLM, SGLang and Transformers recipes, parsers for reasoning and tool calls, tested GPU layouts and a 512K context for five of the six sizes (ev-card-32b, ev-hub).
Safety2/5whether misuse risks are evaluated and guardrails are providedThe models are safety tuned and the report publishes results on four benchmarks in two languages (ev-report-safety).
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted Apache-2.0 weights with no telemetry or claw-back in the licence. Read the two custom Python files that run under trust_remote_code (ev-license, ev-hub).

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
What this means for adoptionYou own the weights, every stage checkpoint, the training and evaluation code and a 23.8B-token data sample, all under Apache-2.0, so you can run and change the whole pipeline from any stage. You do not own a deployable safety case: safety results are the publisher's own, no guard model exists and agent actions are untested. Ownership moves to substantial if an independent evaluation lifts safety to 3 or more.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The 32B card declares license apache-2.0 in its front matter and has no License section.
Model cardread2026-11-19
The 0.9B, 3.7B, 7B, MoVA-36B-A4B and 375B-A23B cards repeat the 32B licence declaration, intended-use text and limitations.
Licenceread2026-11-19
Apache-2.0 is declared in the front matter of all six weights cards and in section 7.1 of the technical report.
Technical_reportread2026-11-19
The report says each Jais 3 model starts from the final post-trained K2 Horizon checkpoint of the same size, keeps its architecture and tokenizer, and is adapted by continual pretraining, SFT, FTPO and RLVR.
Technical_reportread2026-11-19
The report lists released weights, stage checkpoints, training and RL code, a training-data sample, run configurations, W&B logs, audited benchmarks and evaluation code (Table 41).
Technical_reportread2026-11-19
The report names each source with token counts: ArabicTxT360 web text (28.4B), literature collections (9.65B), translated reasoning (30.9B), parallel dialects (26.9B), Arabizi (25.1B) and 11.9B tokens of partner-supplied Arabic books with no named source.
Technical_reportread2026-11-19
The report says GLM-5, Jais 2, Qwen3.8-27B, Qwen3.6-27B, gemma-4-31B-it and MiniMax-M3 generated or judged parts of the training data.
Technical_reportread2026-11-19
Section 6.5 reports publisher-run results on AraSafe (1,254 harmful prompts, above 92% refusal for every size), Arabic Regional Safety, HarmBench and XSTest, scored with gemma-4-26B-A4B-it.
Technical_reportread2026-11-19
The report says Jais 3 leads or matches open models of its size on Arabic benchmarks.
Technical_reportread2026-11-19
Training covers Gulf, Egyptian, Levantine and Darija Arabizi, country knowledge for 22 states, and translation for seven dialects.
Documentationread2026-11-19
All six weights repositories are ungated, hold safetensors only with LFS sha256 values, and report a completed hub scan with no flagged files.
Documentationread2026-11-19
Each weights repository has branches CPT, Cooldown, FTPO and SFT besides main.
Documentationread2026-11-19
The IFM organisation hosts FP8, MLX and GGUF builds of the Jais-3 models, each tagged Apache-2.0.
Documentationread2026-11-19
IFM/Jais-3-Data is an ungated 23.8B-token sample, about 10% of CPT and SFT tokens, in four Arabic subsets licensed ODC-BY or CC BY 4.0.
Documentationread2026-11-19
The xllm, RL360 and gen-eval-arabic repositories are public with Apache-2.0 LICENSE files.
Documentationread2026-11-19
The JaisVoice README dates the Jais 3 release to 2026-10-15.
Model cardread2026-11-19
The K2-Horizon-32B card declares Apache-2.0, says "Fully open" and links code and data.
Technical_reportread2026-11-19
The K2-Horizon abstract describes six fully open models, about 10T synthetic pre-training tokens and a release that follows the LLM360 definition of open source (publisher claims).
Licenceread2026-11-19
The MiniMax Community License grants use for non-commercial purposes and sets conditions on commercial use of the software and its derivative works.
Model cardread2026-11-19
The four Jais-3 Uno adapter repositories carry no licence tag and no licence statement.
Documentationunverified2026-11-19
The ifm.ai blog and tech-report pages return HTTP 403 (Cloudflare challenge), so their text is unavailable.