Ling-3.0-tiny-base

Ling-3.0-tiny-base

Quick answer: Ling-3.0-tiny-base is InclusionAI's (Ant Group's) small-scale pretrain checkpoint from the Ling-3.0 series — a 7.9B-total, 1.3B-activated hybrid-linear MoE model released under MIT license. It shares its training recipe with the larger Ling-3.0-flash-base (validated strategies scale from tiny to flash), scoring 92.2% ARC-C, 80.59% CEval, and 51.83% MMLU-Pro on InclusionAI's self-built base-model eval suite.

At a Glance

Where Ling-3.0-tiny-base leads

• 92.2% ARC-C, 84.01% CCPM — strong knowledge/reasoning for a 1.3B-active-parameter model
• 79.27% HumanEval-Plus, 64.38% MultiPL-E — leads all compared models on these two coding benchmarks
• MIT license — fully open, permissive for commercial use and continued pretraining
• Serves as the validated small-scale testbed for the Ling-3.0 training recipe before scaling to Ling-3.0-flash-base

Where it lags

• Not instruction-tuned: this is a pretrain/WSM-merged checkpoint, not suited for direct chat deployment — see Ling-3.0-tiny for the post-trained instruct model
• 7.20% SimpleQA-Verified — modest factual recall, expected at this parameter scale
• 51.83% MMLU-Pro — trails Qwen3.5-9B-base (57.76%) on graduate-level knowledge

Best for: lightweight continued-pretraining and fine-tuning research, edge/resource-constrained MoE systems research, and rapid experimentation before scaling validated strategies to Ling-3.0-flash-base.

What Ling-3.0-tiny-base Is

Ling-3.0-tiny-base is the smaller of two base-checkpoint sizes in InclusionAI's (Ant Group's open-source arm) newly released Ling-3.0 series. Like its larger sibling Ling-3.0-flash-base, it's released as a full training progression — pretrained-only (Ling-3.0-tiny-base-30T), mid-trained (Ling-3.0-tiny-base-midtrain), and WSM-merged (Ling-3.0-tiny-base, this page) — using Warmup-Stable-and-Merge in place of conventional learning-rate decay, so researchers can explore different decay profiles without rerunning training from scratch.

Architecturally, Ling-3.0-tiny-base uses the same hybrid-linear MoE design as the rest of the series: a highly sparse (1/64) MoE with 128 routed experts (8 activated + 1 shared per token) plus native hybrid linear attention (KDA + Gated MLA), giving it 7.9B total parameters while activating just 1.3B per token. InclusionAI explicitly designed Ling-3.0-tiny-base and Ling-3.0-flash-base to "share the same training recipe," so the tiny checkpoint serves as a fast, low-cost proving ground for training strategies before they're scaled to the 124B flash size. On InclusionAI's self-built eval suite, it's benchmarked against Ling-2.5-mini-base, Qwen3.5-9B-base, and Qwen3.5-4B-base — leading on HumanEval-Plus and MultiPL-E despite its small size.

Note this is explicitly a pretrain/base checkpoint — per InclusionAI's model card, it is "not recommended as-is for direct end-user chat deployment" or "production use without post-training and task-specific validation." For chat/instruct use, see the post-trained Ling-3.0-tiny.

Specs

Pricing

Open weights under MIT license — self-host at no cost via Hugging Face or ModelScope.

Public Benchmark Scores

Scores are self-reported by InclusionAI on the official Hugging Face model card (inclusionAI/Ling-3.0-tiny-base) using InclusionAI's self-built base-model evaluation suite. Only the subset with existing Benchgen benchmark pages is listed above — additional benchmarks from the model card (CCPM, AGIEval, SimpleQA-Verified, CruxEval, MultiPL-E, FullStackBench, LCBench, OlympiadBench, TheoremQA, OmniMath, CommonSenseQA, BBH, LEval) don't yet have Benchgen pages and are flagged as follow-ups below (same set flagged on the sibling Ling-3.0-flash-base page).

Notable Results Not Yet Tracked as Benchgen Benchmarks

• CCPM (84.01%), AGIEval (64.49%), SimpleQA-Verified (7.20% — a distinct, harder-curated variant of OpenAI's SimpleQA, not directly comparable to the existing simpleqa.md page's scores), CruxEval (67.44%), MultiPL-E (64.38%), FullStackBench (39.48%), LCBench (41.59%), OlympiadBench (25.90%), TheoremQA (51.64%), OmniMath (29.70%), CommonSenseQA (83.46%), BBH (80.64%), and LEval (68.37%) are all real, methodology-documented benchmarks per the model card, but don't have existing Benchgen benchmark pages yet — flagged as follow-ups.

Ling-3.0-tiny-base vs Compared Base Checkpoints

Per InclusionAI's model card, Ling-3.0-tiny-base leads all three compared checkpoints on HumanEval-Plus despite having the fewest activated parameters (1.3B) of the group.

FAQ

Is Ling-3.0-tiny-base open source? Yes — MIT license, weights on Hugging Face (inclusionAI/Ling-3.0-tiny-base).

Who makes Ling-3.0-tiny-base? InclusionAI, Ant Group's open-source AI arm.

Is Ling-3.0-tiny-base an instruct/chat model? No — it's a pretrain/WSM-merged base checkpoint, explicitly not recommended for direct chat deployment. Use the post-trained Ling-3.0-tiny for chat/task use.

How big is Ling-3.0-tiny-base? 7.9B total parameters (hybrid-linear MoE), with 1.3B activated per token.

What checkpoints are available for Ling-3.0-tiny? Three stages: Ling-3.0-tiny-base-30T (pretrained only), Ling-3.0-tiny-base-midtrain (mid-trained), and Ling-3.0-tiny-base (WSM-merged, this page) — the last is the recommended starting point for further fine-tuning.

Where can I access Ling-3.0-tiny-base? Self-hosted from the open weights at huggingface.co/inclusionAI/Ling-3.0-tiny-base, also mirrored on ModelScope.

Benchmark scores sourced from the official Hugging Face model card for inclusionAI/Ling-3.0-tiny-base (Aug 2026), InclusionAI's self-built base-model evaluation suite.