Ling 3.0 Flash Guide: 5.1B Active Parameters, MIT Weights

Most cheap models are cheap because they are small. Ling 3.0 Flash is cheap for a different reason: it is a 124B-parameter mixture-of-experts model that only touches about 5.1B parameters per token. That ratio is the reason it exists, and it is also the reason its benchmark numbers look out of proportion to its price.

inclusionAI, the Ant Group lab behind the Ling and Ring families, published the weights under the MIT licence. OpenRouter lists it from 23 July 2026, and it has since become the base for a small family of fine-tunes. This guide covers the base model; the siblings get their own pages.

One active parameter in twenty-four

The model card describes 512 routed experts plus one shared expert, with eight routed experts chosen per token. inclusionAI calls this 1/64 sparsity. The practical result is a forward pass closer in compute to a 5B dense model than to anything with "124B" on the label.

Memory is a different story. All 124B parameters still have to sit somewhere, so the serving footprint is that of a large model even though the per-token compute is that of a small one. The card's SGLang recipe uses four 141GB-class GPUs for the full context, or eight 80GB cards such as the H100. If the distinction is new, active vs total parameters walks through why both numbers matter and for different reasons.

The attention design is the unusual part

Ling 3.0 Flash is a hybrid-linear model. It alternates Kimi Delta Attention layers, a linear-attention variant with gating, with gated Multi-head Latent Attention layers at a 5:1 ratio: 35 KDA layers and 7 MLA layers, plus two dense layers. inclusionAI says the hybrid design was used from the start of pretraining rather than retrofitted.

Why it matters: linear attention layers do not grow their cost with the square of the sequence length, so a model built mostly from them is cheaper to run over long inputs. The full-attention MLA layers are there to keep precise retrieval across the context working. The context window is 262,144 tokens, reached through a training schedule that stepped from 8K to 32K to 256K.

The honest caveat is that hybrid-linear designs are newer than plain transformers, and the tooling is still catching up. The card points at a dedicated SGLang image and a vLLM build from source, with a ling3 reasoning parser and tool-call parser. Expect to follow the recipe rather than drop it into whatever inference stack you already run.

What inclusionAI claims

The scores the card states in text are:

• SWE-bench Pro: 56.6
• SWE-bench Multilingual: 72.4 resolved
• AIME 2026 (MathArena): 93.2
• HMMT February 2026 (MathArena): 87
• Humanity's Last Exam: 22.7

The card also says the model matches or beats Ring-2.6-1T, the previous trillion-parameter flagship, on key benchmarks while using about 12.4% of its total parameters and 8.1% of its active ones. The later VL model card puts the base at 38 on the Artificial Analysis Intelligence Index v4.1.1. Several agentic results, including Terminal-Bench 2.1 and MCP-Atlas, appear only as charts, so they are left out here.

Two cautions. These are vendor numbers on vendor harnesses. And the gap between SWE-bench Pro and Humanity's Last Exam is telling: the model is tuned for agentic coding and maths, not for broad expert knowledge. A small active-parameter count limits how much the model can know, even when training teaches it to use tools well.

Thinking is on by default, and that changes the bill

Ling 3.0 Flash ships with reasoning enabled in both the chat template and the parser. On SGLang you can switch it off per request with "chat_template_kwargs": {"enable_thinking": false}. The card recommends temperature 0.6, top_p 0.95 and top_k 20.

The per-token price is tiny. Our OpenRouter catalogue sync listed it at about $0.021 per million input tokens and $0.063 per million output at the time of writing, with a 32,768-token output cap on that endpoint. But a reasoning model spends output tokens you never see, and a small model can be verbose about it. The token cost of reasoning models explains why price per token and price per answer drift apart. Measure the second one.

Which Ling to pick

The base model now has four relatives, and the choice is mostly about the job:

• Ling 3.0 Flash: general coding, agent loops, maths, bulk text work where cost per call dominates.
• Ling 3.0 Flash Fin: the same architecture with continued training on financial data, aimed at research and filings work.
• Ling 3.0 Flash Sante: the health and medicine variant, built on the same base.
• Ling 3.0 Flash VL: adds image and video input; its OpenRouter listing describes it as 5.5B active rather than 5.1B.
• Ling 3.1 Flash: a much larger successor, 560B total and 25B active, for when the base model runs out of headroom.

A sensible way to adopt it

Do not swap it in as your only model. Its natural role is the cheap first tier: classification, extraction, short code edits, first-pass agent steps. Run it on twenty real tasks with thinking on and off, and record both pass rate and output tokens per task. If it clears the bar on the easy half, route that half to it and escalate the rest. That pattern is covered in model routing and fallbacks, and with a model this cheap the savings from getting the routing right are large.