Vibeleaderboard
← All Intel
Intel / post

Ling-3.0-tiny open weights with benchmarks and local throughput

Source
x.com
Date
Ant Ling@AntLingAGI

Ling-3.0-tiny is now available as an open-weight model in BF16, FP8 and INT4. On Artificial Analysis, it scores 25 on the Intelligence Index and 16 on the Agentic Index, with 772 Elo on GDPval-AA v2 and 20.80 on τ³-Banking—built for real task execution. 🧵

Context

Ant Group's Ling team made Ling-3.0-tiny downloadable on August 10, 2026, four days after it announced the model, and its release post offers the weights in BF16, FP8 and INT4. Those are storage formats for the model's numbers: BF16 uses 16 bits per weight, FP8 uses 8 and INT4 uses 4, so the smaller formats need less memory to load and run. The model is a . Its model card describes 128 routed expert sub-networks, of which 8 plus one shared expert handle each , so about 1.3 billion of its 7.9 billion parameters do the work for each token. The weights carry an MIT license.

The figures Ant quotes come from Artificial Analysis, an independent firm that runs its own evaluations: 25 on its Intelligence Index, 16 on its Agentic Index, a 772 Elo rating on GDPval-AA v2 and 20.80 on τ³-Banking. Artificial Analysis's own post reports the same Intelligence Index score and puts it beside gpt-oss-120b (high) at 24, a model with 15 times as many total parameters. The same post calls the model's token usage high: it produced 213 million output tokens to complete the index, close to the 240 million used by the much larger Ling 3.0 Flash. A small model that reasons at length is cheap per token but not necessarily quick per answer, and the model card lists about 18 seconds of end-to-end latency for a 500-token response, reasoning included.

The local speed numbers are Ant's own. Its model card reports that with FP8 weights the model reaches about 100 to 105 tokens per second on an NVIDIA DGX Spark and 86 to 90 tokens per second on an M4 Pro MacBook, without stating batch size or how many requests ran at once. Artificial Analysis describes the model as small enough to run locally in many settings; the caveat is that its answers take more tokens than its size suggests.

Terms in this piece · Glossary
  • mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
  • open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
More from Ant Ling
Recommended reads
Comments

Checking sign-in…

Loading comments…