Jais 3 Lands in vLLM Before It Lands in the Press Release

Check the date on the J​ais-3-375B-A23B serving recipe and it does not say what the launch coverage says. The vLLM recipe page for the flagship of the J​ais 3 family carries a last-updated stamp of 5 October 2026. The weights for J​ais-3-32B, J​ais-3-7B, J​ais-3-3.7B, J​ais-3-0.9B and J​ais-3-MoVA-36B-A4B went up on the Hub on 29 September. The press release from the Institute of Foundation Models and Inception42 is dated 15 October. That ordering is the actual news here, and it is a better story than the one the announcement tells: the serving stack did not scramble to catch up with an Arabic frontier model — it was ready ten days early. This piece is about what that readiness looks like in practice, what the family ships, and the two questions nobody has answered yet.

The recipes are real, and they are specific

Framework support announcements are usually a sentence and a link. This one has executable content, which is worth reading closely because it tells you what the deployment actually costs before you download 700-odd gigabytes of weights.

• vLLM — the recipes target vllm>=0.30.0/. The 375B flagship wants --tensor-parallel-size 16 --enable-expert-parallel/ on a bfloat16 checkpoint, which is 16× H100 80GB across two 8-GPU nodes in IFM's own reference deployment.

• Smaller siblings get smaller footprints — J​ais-3-32B serves at --tensor-parallel-size 2/, J​ais-3-7B and J​ais-3-3.7B at --tensor-parallel-size 1/. J​ais-3-MoVA-36B-A4B wants TP=2 with expert parallelism enabled, and its 4B active parameters per token are the reason it does not need a node farm.

• SGLang — the parallel recipe targets sglang>=0.5.20/, with the two-node 375B configuration running TP=16 and EP=16, FlashAttention-3, and multithreaded weight loading explicitly disabled.

• H200 changes the arithmetic — at 141GB per GPU the flagship fits on a single 8-GPU H200 node. IFM's card says to drop tensor and expert parallelism to 8 in that case. One node instead of two is not a rounding difference in a cost model.

• AMD is validated too — the recipe page carries an MI355X configuration verified on an 8-GPU gfx950 node, using AITER for the fused MoE and the unified attention backend, with VLLM_ROCM_USE_AITER=1/. A day-zero N​VIDIA-only story is common; a day-zero story with a second vendor's silicon verified is not.

What makes the support more than a compatibility checkbox is that the architecture is genuinely custom. The Hub configs declare model_type: k2_horizon/ and K2HorizonForCausalLM/ with custom remote code, the 375B carrying 192 experts and routing 8 per token. That is the kind of model that normally arrives on a framework weeks after launch with a patch list in the release notes, not with a checked-in recipe dated before the press release.

Arabic-native agentic, in the flags it actually takes

"Agentic" is the word in the announcement headline, and the serving commands are where you can test whether it means anything. It mostly does.

• --reasoning-parser k2_horizon/ separates the thinking trace from the answer. Thinking is returned in reasoning_content/ and the answer in content/, which is the difference between a chain of thought you can log and one you have to regex out of the response body.

• --tool-call-parser k2_horizon/ with --enable-auto-tool-choice/ turns on tool use, and the model supports three tool-call formats — json/, xml/ and xml_typed/ — selectable per request through chat_template_kwargs/, defaulting to xml/.

• Reasoning depth is a per-request setting, not a deployment decision. {"reasoning_effort": "high"}/ is the default and what every published number was produced with; "low"/ turns thinking off entirely, and IFM's own card says low is not recommended outside research on reasoning effort.

• Sampling parameters swap by language, which is unusual and worth catching. Default is temperature 0.2, top_p 0.95, top_k 20. In thinking mode IFM asks for temperature 1.0 for English and 0.7 for Arabic. A single global temperature setting is the wrong configuration for this model.

• Output budget is large. The card recommends at least 32,768 output tokens — a strong hint about how long the reasoning traces run, and a real constraint on your per-request timeout if you are used to 4K completions.

What the family actually ships

Six base checkpoints, not one. The spread matters more than the flagship, because it determines whether the family is a research artifact or something you can put behind a workload.

• J​ais-3-0.9B — 0.9B dense, 128K context, the only member that is not 512K. RoPE scaling via YaRN is in the vLLM command, so the long-context behaviour on this one is an explicit override rather than a trained default.

• J​ais-3-3.7B, J​ais-3-7B, J​ais-3-32B — dense, all at up to 524,288 tokens of context, described by IFM as small, medium and large dense members.

• J​ais-3-MoVA-36B-A4B — a mixture-of-experts model with Mixture-of-Values attention. The recipe page puts it at 37.44B stored parameters including embeddings and 5.95B active per token, against the card's headline 36B and 4B. Both numbers are the vendor's; they count embeddings differently. Take the recipe page's figures when you are sizing GPUs and the card's when you are quoting the model.

• J​ais-3-375B-A23B — the flagship, 375B stored and 23B active by the card, 379.17B and 26.67B by the recipe page. Same counting discrepancy, same resolution.

• Quantised and edge formats are already out — FP8 builds, GGUF for llama.cpp and Ollama, and MLX 4-bit, 8-bit and bf16 for Apple silicon, published across the IFM organisation. The MLX builds for the 32B landed on 1 October, the GGUF builds for the 3.7B and 7B on 13 October. Nobody was waiting on a follow-up release to run these locally.

• Licence is Apache 2.0 across the base family, with no gating on the inception42 repositories. That is a materially different posture from several of the Arabic models this family now competes with.

There is also J​ais-ASR in the same organisation — a speech-recognition model built on a J​ais-3-4B backbone, covering Modern Standard Arabic plus dialects from 19 Arab countries with code-switching into English, French, Spanish and Italian, and reporting 22.85% WER and 8.37% CER on the Open Universal Arabic ASR Leaderboard. It is a separate model with its own licence file and its own vLLM plugin, but it is the reason the announcement says "speech" alongside "writing and agents".

The numbers are the vendor's, and this vendor audits them

Every benchmark figure below comes from IFM's own model cards and has not been reproduced by a third party. That sentence needs to be in front of the numbers, not after them.

• Arabic MMLU, flagship: 89.3. Arabic MMLU with diacritics: 88.2.

• Belebele: 95.6. Dialectal Arabic MMLU: 89.6. Dialect identification (QADI): 71.1. Alyah: 87.6.

• AraSTEM: 95.5. Arabic GSM8K: 86.8. GATmath: 98.4.

• Arabic IFBench, prompt-level strict: 93.9. Arabic IFEval: 77.7. FLORES200+ LLM-as-a-judge: 91.9.

• Agentic: Terminal Bench 2.1 at 73.8, GDPVal at 73.2, MCPMark at 68.5, SWE-Bench Verified (offline) at 70.2, GAIA v2 lit (Arabic) at 50.3. Apex-Agents is the one that keeps the sheet honest at 18.6.

The methodology note is the part worth crediting. IFM marks benchmarks that have been audited for sample quality, with ambiguous questions, invalid answer options and indefensible gold labels excluded, and it publishes the audit methodology, the excluded samples and the verified benchmark list as a separate collection. This is not peer review and it does not make a vendor score independent, but a lab that tells you which of its own numbers it cleaned is a lab you can read more confidently than one that does not. Scores on the cleaned sets are also not the same thing as scores on the benchmark as commonly run, and the starred marks are the ones to check before you compare against somebody else's sheet.

Two things nobody has answered

There is no pricing. Not on the Hub cards, not on the recipe pages, and the vendor's own blog is behind a bot challenge that no automated reader gets through. There is no published per-token rate for any member of this family to compare against anything. Anyone quoting a price for J​ais 3 today is quoting something they invented, or quoting what a third-party host happens to charge for running it on their own GPUs — which is a hosting price, not a model price.

There is no independent evaluation. Artificial Analysis has no entry, no arena has posted head-to-head results, and no neutral leaderboard has run the family. The technical report is listed on the model card as "coming soon", and so is the code repository. Until something outside the lab runs these weights, the correct reading of a 95.5 on AraSTEM is "IFM says 95.5".

If you want to try it before there is a bill

The weights are open and Apache 2.0, so the cheapest evaluation is your own hardware — and the recipes above are short enough to run today. What is harder is putting the family behind something real, because you would be standing up a 16-GPU cluster, or renting one, to find out whether a model with no independent score belongs in a production path.

That is the case where a routing layer earns its keep rather than decorating an architecture diagram. OrcaRouter gives you one API across 200-plus models with automatic failover, so an Arabic workload can send a slice of traffic at an unproven model and let something already measured absorb the rest — the evaluation happens against real prompts, and a bad response or an unhealthy endpoint fails over instead of taking the request down with it. The catalogue is worth checking before you plan around a self-host, because it moves. If a model is not on it today it may be next week, and the pricing page is the honest source for what anything costs once it is.

What to watch next is narrow and specific: the technical report, the code repository, a first number from someone who is not IFM, and a published price. Any one of those changes this piece. None of them has landed yet, which is exactly why the day-zero framework support is the part of this launch that is currently verifiable.

What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.