Jais 3 goes open, Step 5 and Ling 3.1 land, Drama 3 speaks

No single frontier launch today, but the open-weights shelf got heavier. I read four model cards, and three of them are sparse mixture-of-experts models north of 350 billion parameters.

Jais 3: an Arabic and English flagship under Apache 2.0

Inception published the Jais-3 family, and the card for Jais-3-375B-A23B is the one to read. It stores 375B parameters, runs 23B per token, reasons before it answers and handles a 512K context window. The license is Apache 2.0, and the language tags are Arabic and English.

The card compares it against Qwen3.5-397B-A17B, MiniMax M3, Qwen3.8-Flash-Next and NVIDIA Nemotron 3 Ultra, with Arabic open-ended generation sitting beside the usual reasoning suites. The authors say they audited some benchmarks for sample quality and report scores on the cleaned sets, and that a few benchmarks are still to come. I have not rerun any of it, so I treat the table as the lab's own claim.

The reference deployment is 16 H100s across two nodes, or one 8-GPU H200 node, with vLLM and SGLang recipes. The family also includes smaller checkpoints, from 0.9B up to a 36B variant, so there is a way in without a cluster.

Step 5 Preview and Ling-3.1-flash: two big agent models with 1M context

StepFun calls Step 5 Preview its flagship for agentic work. It is a 600B mixture-of-experts model with 27B active, takes text, images and video, and reaches 1M tokens of context. The card stresses repository-level software engineering, long tool-using runs and finance. FP8 fits one H200 node, and a vLLM image is published for it. The license is custom, so read it before you build on it.

inclusionAI's Ling-3.1-flash is 560B with 25B active, a hybrid-linear design with 63 linear-attention layers and 9 gated MLA layers, and a 1M-token window. It is MIT licensed. The card lists results on Terminal-Bench 4.0, SWE-bench Pro and several office and finance suites, and it mentions a coding harness the team may open-source later. The SGLang image ships on day zero with multi-token-prediction decoding.

Both cards read as agent-first. The pitch is long runs with many tool calls, not a single clever answer.

Fish Audio Drama 3: one script, up to 16 voices

Drama 3 is a text-to-speech model with 35B parameters and 3B active, built for multi-character scenes and long-form audio. One text can carry up to 16 speakers through tags, long input is chunked and streamed, and the card covers voice cloning and editing a take by sending back the previous audio tokens. Inline tags such as sighs and laughs land at the position you write them.

It runs on SGLang-Omni, and the license is Fish Audio's own community license, not an open-source one. For anyone building audio dramas or agent voices, this is worth a listen.

What I'd watch next

Independent Arabic evals for Jais 3 against the Qwen and MiniMax models it cites. Whether Step 5 moves from Preview to a final release. And whether the 1M-context claims on Step 5 and Ling hold up in real agent runs, not only on the cards.