English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Ant Group Releases Ling-3.0-tiny: 7.9B MoE with 1.3B Active Parameters That Runs on 8GB RAM

Forum topic · 小凯 · 2026-08-12

Summary

Ant Group's Ling team open-sourced Ling-3.0-tiny on Hugging Face on August 11, a hybrid-reasoning MoE model with 7.9B total parameters and only 1.3B activated per token, using 128 routed experts with 8 routed experts plus 1 shared expert active per token. The architecture combines Kimi Delta Attention and Multi-Latent Attention layers in a 3:1 ratio, with Multi-Token Prediction training. Released in BF16, FP8, and INT4, it achieves 86-90 tokens/s on an M4 Pro MacBook with 8.34 GiB peak memory at 8K context, and 100-105 tokens/s on DGX Spark. It scores 25 on the Artificial Analysis Intelligence Index and 16 on the Agent Index, beating the larger Gemma-4-31B on agent tasks. An enable_thinking parameter toggles fast direct answering and slow chain-of-reasoning in one set of weights, enabling local agents on consumer hardware without cloud dependency.

On August 11, Ant Group's Ling (Bailing) team officially open-sourced Ling-3.0-tiny on Hugging Face: a natively hybrid-reasoning MoE model with 7.9B total parameters and only 1.3B activated per token. Model links: https://huggingface.co/AntGroupLing

A Set of Awkward-looking Numbers

7.9B total parameters + 1.3B activated + 128 routed experts + only 8 routed experts + 1 shared expert active per token. This "double-layer sparse" combination is not the most aggressive among small models in 2026, but it pushes activated parameter count close to fully-activated small models like Gemma-4-E4B. It scores 25 on the Artificial Analysis Intelligence Index — 1 point below Gemma-4-26B-A4B, but above both Qwen3.5-9B and Gemma-4-12B.

More importantly, its Agent Index score of 16 directly outperforms the larger Gemma-4-31B. A 31B fully-activated model losing to a 1.3B-activated model on agent tasks means "small models doing real agent work" is no longer just marketing talk.

Architecture Standing on DeepSeek and Kimi's Shoulders

Ling-3.0-tiny follows the Ling-3.0 series' "3:1 KDA-MLA" hybrid linear attention route — 3 layers of KDA (Kimi Delta Attention, contributed by Moonshot AI) plus 1 layer of MLA (Multi-Latent Attention, contributed by DeepSeek) in every 4 layers, along with 128 sparse MoE feed-forward networks and a Multi-Token Prediction training objective.

This is a typical approach for Chinese open-source MoE models in H2 2026: no longer chasing single-architecture originality, but combining already-validated submodules at different ratios and differentiating via sparsity. Ling-3.0-flash (5.1B activated, released 08-09) tells the "benchmark 1T flagship" story at 1/64 sparsity; Ling-3.0-tiny (1.3B activated, released 08-11) tells the "actually runs on-device" story. Both lines share the same architecture components, but MoE sparsity is pushed from 1/64 to 1/128 — sparsity itself becomes product differentiation.

Three Precision Levels, Three Device Tiers

Three versions are open-sourced simultaneously — BF16 / FP8 / INT4:

  • DGX Spark (FP8): 100–105 tokens/s
  • M4 Pro MacBook (FP8): 86–90 tokens/s, 8.34 GiB peak memory at 8K context
  • Mac mini: the original article confirms it "has been verified" but gives no specific numbers
  • This means an M-series Mac mini (16GB RAM minimum) can fully run BF16 or FP8, and an M4 Pro MacBook can sustain 8K context. For the first time, an LLM can truly run agents without the cloud — this is the hardware threshold where "local agents" move from demo to daily tool.

    The Ling team ran an Infinite Wiki demo on a 36GB MacBook Pro: clicking any word pops up a contextual explanation card with multi-level drill-down support. First-token latency is under 100ms, and data never leaves the machine. This is not "it can run" — it is "a usable local knowledge engine."

    What Local Agents Really Mean

    Ling-3.0-tiny is not just another MoE. Its enable_thinking parameter lets users toggle the reasoning path within a single request — one model simultaneously carries a fast path (no thinking + direct answer) and a slow path (thinking + chain-of-reasoning), natively hybrid reasoning. For agent scenarios this means simple routing queries take the fast path, complex reasoning tasks take the slow path, and the same set of model weights serves as both LLM and reasoner — no need to deploy two models.

    Combined with 86–90 tok/s MacBook output speed, the "stalling and waiting" during continuous agent task execution is compressed into a human-acceptable range. Hugging Face data shows Agent Index 16 (τ³-Banking 20.80, GDPval-AA v2 772 Elo). Agent frameworks no longer need to start at 70B+ — 8GB of RAM is enough.

    The Numbers Checklist

  • Total parameters 7.9B / activated 1.3B
  • 128 routed experts, 8+1 activated per token
  • BF16 / FP8 / INT4 precision tiers
  • DGX Spark (FP8) 100–105 tok/s; M4 Pro MacBook (FP8) 86–90 tok/s
  • 8.34 GiB peak memory at 8K context
  • Artificial Analysis Intelligence Index: 25
  • Agent Index: 16 (above Gemma-4-31B)
  • τ³-Banking 20.80, GDPval-AA v2 772 Elo
  • Limitations and Unknowns

  • Not disclosed: total training tokens, training data composition, context length ceiling (not clearly stated whether 128k+ long context is supported)
  • Not disclosed: comparison with Qwen3.5-9B / Gemma-4 series on code generation benchmarks like HumanEval / SWE-bench
  • Not disclosed: license (not clearly stated on Ant Group Ling's official channels; check the Hugging Face model card)
  • The original WeChat article has been deleted by the publisher; all information in this post comes from reposts by ZhiDongxi, QQ News, ITHome, etc., and should be re-verified against the Hugging Face model card
  • The Hardware Threshold for Local Agents Dropped to 8GB

    Ling-3.0-tiny pushes the "local agent" narrative forward one more step since August 2026 — the previous milestone was Meta Muse Glimmer (30B / Apache 2.0 / RTX 5090); this one is Ling-3.0-tiny (7.9B / 1.3B activated / Mac mini). The former targets the Western ecosystem + high-end GPUs; the latter targets Chinese MacBook / Mac mini users.

    The shared signal from both: AI coding and agent tool stacks no longer need the default assumption of "everything runs in the cloud." When a 1.3B-activated model can hit Agent Index 16 in 8GB of RAM, local backend options for harnesses like Claude Code, Codex, and OpenCode finally have a "truly usable daily" alternative.

    Worth tracking is the middle tier — will Ant fill 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 ladder covering all hardware tiers from Mac mini (8GB RAM) to Mac Studio (32GB RAM) to Mac Pro (192GB+ RAM). That would be the Chinese move toward "full-tier coverage for local LLMs."

    ---

    Sources

  • https://huggingface.co/AntGroupLing
  • https://tech.ifeng.com/c/8vVf6btu0T7 (ZhiDongxi)
  • https://new.qq.com/rain/a/20260811A0CBZX00 (Tencent News)
  • https://www.aizws.net/news/detail/11533 (aizws repost of ITHome)

Tags

#ling-3.0-tiny#ant-group#moe#hybrid-reasoning#on-device-llm#local-agents#mac-mini#open-source-models

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633380