Ling-3.0-flash-base

Ling-3.0-flash-base

Quick answer: Ling-3.0-flash-base is InclusionAI's (Ant Group's) merged/WSM pretrain checkpoint from the Ling-3.0 series — a 124B-total, 5.1B-activated hybrid-linear MoE model released under MIT license. On InclusionAI's self-built base-model eval suite it scores 90.0% CCPM, 89.76% CEval, 67.74% MMLU-Pro, and 79.0% MATH500, outperforming comparably-sized base checkpoints like Ling-2.5-flash-base and NVIDIA-Nemotron-3-Super-base on most reported benchmarks despite activating the fewest parameters of the compared models (5.1B vs 7.4B–21B).

At a Glance

Where Ling-3.0-flash-base leads

• 90.0% CCPM, 96.27% ARC-C — strong knowledge/reasoning for its active-parameter budget
• 89.76% CEval, 67.74% MMLU-Pro — competitive general-knowledge scores versus larger-active-param peers
• 81.10% HumanEval-Plus, 70.79% MultiPL-E — strong base-model coding ability
• 79.0% MATH500 — leads all compared models on this math benchmark
• MIT license — fully open, permissive for commercial use and continued pretraining
• Only 5.1B activated parameters — most efficient of the compared checkpoints (peers range 7.4B–21B active)

Where it lags

• Not instruction-tuned: this is a pretrain/WSM-merged checkpoint, not suited for direct chat deployment — see Ling-3.0-flash for the post-trained instruct model
• 47.89% OlympiadBench — trails NVIDIA-Nemotron-3-Super-base (54.52%) on olympiad-level math
• 52.62% LongBench — solid but not class-leading on long-context comprehension versus some peers

Best for: continued pretraining, mid-training, domain-adaptation SFT, and RL post-training research where a strong, efficient (low active-parameter) MoE base checkpoint with a documented training recipe is needed.

What Ling-3.0-flash-base Is

Ling-3.0-flash-base is the larger of two base-checkpoint sizes in InclusionAI's (Ant Group's open-source arm) newly released Ling-3.0 series, InclusionAI's "most efficient language foundation model family to date." Rather than releasing a single checkpoint, InclusionAI published the full training progression for each size — pretrained-only, mid-trained, and WSM-merged ("Warmup-Stable-and-Merge," which replaces conventional learning-rate decay with weighted checkpoint merging) — to support research on continued pretraining and offline exploration of different decay strategies. Ling-3.0-flash-base refers to the final WSM-merged checkpoint, the recommended starting point for further fine-tuning.

Architecturally, Ling-3.0-flash-base is a highly sparse (1/64) hybrid-linear MoE: 512 routed experts (8 activated + 1 shared per token) combined with a native hybrid linear attention design (KDA + Gated MLA) for efficient long-context processing, adopted from the start of pretraining rather than retrofitted. This gives it 124B total parameters while activating only 5.1B (non-embedding) per token. On InclusionAI's self-built evaluation suite spanning knowledge, code, math, reasoning, and long-context domains, it's benchmarked directly against Ling-2.5-flash-base, NVIDIA-Nemotron-3-Super-base, Step-3.5-Flash-base, Hy3-preview-base, and MiMo-V2.5-base — outperforming all of them on several benchmarks (CCPM, MMLU-Pro, HumanEval-Plus, MultiPL-E, BigCodeBench, MATH500, CommonSenseQA, BBH, LongBench, LEval) despite activating the fewest parameters per token of the group.

Note this is explicitly a pretrain/base checkpoint — InclusionAI's model card states 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-flash.

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-flash-base) using InclusionAI's self-built base-model evaluation suite. Because these are traditional academic base-model benchmarks (as opposed to the agentic/tool-use benchmarks typical of instruct-model releases), only the subset with existing Benchgen benchmark pages is listed above — several 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.

Notable Results Not Yet Tracked as Benchgen Benchmarks

• CCPM (90.00%, Chinese classical poetry), AGIEval (77.50%), SimpleQA-Verified (26.90% — a distinct, harder-curated variant of OpenAI's SimpleQA, not directly comparable to the existing simpleqa.md page's scores), CruxEval (78.50%), MultiPL-E (70.79%), FullStackBench (51.16%), LCBench (57.76%), OlympiadBench (47.89%), TheoremQA (61.10%), OmniMath (47.65%), CommonSenseQA (89.93%), BBH (89.17%), and LEval (83.24%) 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-flash-base vs Compared Base Checkpoints

Per InclusionAI's model card, Ling-3.0-flash-base outperforms all five compared base checkpoints on MATH500, HumanEval-Plus, and MMLU-Pro despite having the smallest activated-parameter count (5.1B) of the group.

FAQ

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

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

Is Ling-3.0-flash-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-flash for chat/task use.

How big is Ling-3.0-flash-base? 124B total parameters (hybrid-linear MoE), with 5.1B (non-embedding) activated per token.

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

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

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