Jais 2 ships: a 70B open Arabic LLM at 2,000 tokens per second
Two stories today show where the field's energy is going: open models getting culturally specific, and a hard, numbers-backed look at the "skills" layer everyone is bolting onto their agents.
Jais 2 is here: a 70B open-weight Arabic model that serves 2,000 tokens a second. Inception (a G42 company), Cerebras, and MBZUAI released Jais 2, which they describe as the largest open Arabic-centric LLM trained from scratch, alongside an 8B variant. It ships as a chat app on the web, iOS, and Android, powered by Cerebras hardware the authors say delivers up to 2,000 tokens per second — fast enough that Arabic-speaking users get near-instant replies. The models top the usual English-centric benchmarks on Arabic tasks (OALL2, AraGen) and hold their own on culturally grounded ones like poetry, religion, cuisine, and dream interpretation, and they're released under a commercially permissive license on Hugging Face.
Why it matters: most open-weight momentum has been English-first, and Arabic — spoken by more than 400 million people — has lagged behind. A strong, permissively licensed Arabic model is a building block for sovereign-AI efforts across the Middle East and for any product that needs to actually understand the language and its culture rather than just transliterate it. The Cerebras speed angle is the part that makes Jais 2 usable in production instead of a research artifact.
My take: "open-weight, culturally tuned, and fast" is the combination that turns a model release into something developers adopt. If the 2,000-tokens-per-second figure holds in real deployments, Jais 2 becomes the default for Arabic AI — and a template other language communities will copy.
A new study says LLM "skills" mostly work by steadying the agent, not by teaching it facts. Researchers ran controlled experiments across benchmarks, harnesses, and models and found that skills help when noisy past trajectories become procedural anchors that stabilize what the agent does next. Procedural anchoring accounted for 65.7% of skill cases versus just 4.5% for explicit knowledge injection — skills are scaffolding for behavior, not a knowledge dump. The catch: retrieval is a real bottleneck, and as the skill pool grows from 5 to 100, actual-use precision drops from 29.6% to 3.3%.
Why it matters: every coding-agent and agent platform is racing to ship skills marketplaces, but this paper argues most of the benefit is about keeping the agent on rails — and that benefit collapses once you have too many skills to find the right one. It's a caution sign for teams stuffing hundreds of skills into a context window and assuming more is better.
My take: treat skills as guardrails, not a knowledge base, and keep the pool small enough that retrieval actually works — or the whole mechanism quietly stops helping.
What to watch: the 8B Jais 2 variant is the one to watch for on-device and edge Arabic deployments. If it holds up, the smaller model may matter more than the headline 70B.
Do you think "skills" for AI agents are overhyped, or just misunderstood? Tell us in the comments.
Sources: arXiv — Demystifying Agent Skills · arXiv — Jais 2 · Cerebras — Jais 2 release · MBZUAI — Jais 2 announcement · WAM — Inception, Cerebras, MBZUAI release Jais 2
