Ant Group (Ant Ling) officially open-sourced Ling-3.0-tiny on Hugging Face on August 11. It is a native hybrid reasoning MoE model with 7.9B total parameters and only 1.3B activated per token. Model page: https://huggingface.co/AntGroupLing
A Counterintuitive Numbers Profile
7.9B total + 1.3B activated + 128 routed experts + 8 routed experts active per token + 1 shared expert. This "two-tier sparsity" combination is not the most aggressive in 2026's small-model lineup, but it pushes activated parameters down to a level comparable to fully-activated small models such as Gemma-4-E4B.
On the Artificial Analysis Intelligence Index it scores 25, one point below Gemma-4-26B-A4B and higher than Qwen3.5-9B and Gemma-4-12B.
More importantly, its Agent Index of 16 surpasses Gemma-4-31B — a fully-activated 31B model loses to a 1.3B-activated one on agent metrics. "Small models running large agents" is no longer a marketing slogan.
An Architecture Standing on DeepSeek and Kimi's Shoulders
Ling-3.0-tiny continues the Ling-3.0 series' "3:1 KDA-MLA" hybrid linear attention route — every 4 layers contain 3 layers of KDA (Kimi Delta Attention, contributed by Moonshot) and 1 layer of MLA (Multi-Latent Attention, contributed by DeepSeek). The 128-sparse MoE feed-forward network and the Multi-Token Prediction training objective follow the same playbook.
This is a typical 2026 H2 path for Chinese open-source MoE: rather than pursuing single-architecture originality, validated sub-modules are combined in fixed ratios and product differentiation comes from the degree of sparsity. Ling-3.0-flash (5.1B activated, released 08-09) carries the "1/64 sparse, 1T-class benchmark" narrative; Ling-3.0-tiny (1.3B activated, 08-11) carries the "actually runs on-device" narrative. Both share the same architectural components; only MoE sparsity moves from 1/64 to 1/128.
Three Precision Tiers, Three Device Tiers
Three checkpoints ship together: BF16 / FP8 / INT4.
- DGX Spark (FP8): 100–105 tokens/s
- M4 Pro MacBook (FP8): 86–90 tokens/s, 8K context peak memory 8.34 GiB
- Mac mini: officially "validated" but no concrete number given
- 7.9B total / 1.3B activated
- 128 routed experts, 8+1 active per token
- BF16 / FP8 / INT4 three precision tiers
- DGX Spark (FP8) 100–105 tok/s, M4 Pro MacBook (FP8) 86–90 tok/s
- 8K context peak memory 8.34 GiB
- Artificial Analysis Intelligence Index 25
- Agent Index 16 (higher than Gemma-4-31B)
- τ³-Banking 20.80, GDPval-AA v2 772 Elo
- Not disclosed: total training tokens, training data composition, context-length ceiling (no explicit statement of 128k+ long-context support)
- Not disclosed: code-generation comparison numbers against Qwen3.5-9B / Gemma-4 series on HumanEval / SWE-bench
- Not disclosed: license (not explicitly stated on Ant Ling's official channels; must be checked on the Hugging Face model card)
- Original WeChat article deleted by publisher; current information comes from third-party reprints (Zhidongxi, QQ News, IT Home) and needs cross-checking against the Hugging Face model card
- https://huggingface.co/AntGroupLing
- https://tech.ifeng.com/c/8vVf6btu0T7
- https://new.qq.com/rain/a/20260811A0CBZX00
- https://www.aizws.net/news/detail/11533
In practical terms: a base M-series Mac mini (16GB RAM and up) can fully run BF16 or FP8; an M4 Pro MacBook holds a stable 8K context. For the first time an LLM genuinely does not need the cloud to run an agent — the hardware threshold at which "local agents" moves from demo to daily tool has been crossed.
Ant Ling also demonstrated an "Infinite Wiki" demo on a 36GB MacBook Pro: clicking a word pops up a contextual explanation card with multi-level drill-down, first-token response under 100 ms, and no data leaves the device. This is not "runs" — it is "a usable local knowledge engine."
What "Local Agent" Really Means
Ling-3.0-tiny is not just another MoE. Its enable_thinking parameter lets users toggle the reasoning path within a single request — the same model carries both a fast path (no thinking, direct answer) and a slow path (thinking + chain-of-thought), making it a native hybrid reasoner. For agents this means simple routing queries take the fast path, complex reasoning tasks take the slow path, and the same model weights simultaneously serve as both LLM and reasoner — no need to deploy two copies.
Combined with 86–90 tok/s on MacBook, the "stutter-wait" between consecutive agent actions drops into a human-acceptable range. With an Agent Index of 16 (τ³-Banking 20.80, GDPval-AA v2 772 Elo) on Hugging Face, agent frameworks no longer require a 70B+ starting point — 8GB of memory suffices.
Spec Sheet
Limitations and Unknowns
The Local-Agent Hardware Threshold Drops to 8GB
Ling-3.0-tiny pushes the "local agent" narrative of August 2026 one step further. The previous stop was Meta Muse Glimmer (30B / Apache 2.0 / RTX 5090); this stop is Ling-3.0-tiny (7.9B / 1.3B activated / Mac mini). The former targets the Western ecosystem and high-end GPUs, the latter targets Chinese MacBook / Mac mini users.
The shared signal: AI coding and agent toolchains no longer assume a "everything in the cloud" default. When a 1.3B-activated model can run an Agent Index of 16 in 8GB of memory, harnesses like Claude Code, Codex, and OpenCode gain a viable local backend for daily use.
A development worth watching is the middle tier — whether Ant Group fills the gap between Ling-3.0-tiny (1.3B activated) and Ling-3.0-flash (5.1B activated, released 08-09) with a 3B-activated model. If so, Chinese on-device MoE would have a 1B / 3B / 5B staircase covering Mac mini (8GB), Mac Studio (32GB), and Mac Pro (192GB+). That would be a fully-tiered domestic local-LLM lineup.
---
Sources