Ling 3.0 Tiny vs LFM2.5-2.6B-Base: Two Kinds of Free, and Only One Answers You Today

Both of these models are free. That is almost the only thing they have in common. Ling 3.0 Tiny, Ant Group InclusionAI's 7.9-billion-parameter mixture-of-experts reasoning model, is free because someone is running a server and pointing an API at it — you paste a key into a client and it answers your paragraph right now, no training run, no GPU. LFM2.5-2.6B-Base, Liquid AI's 2.69-billion-parameter pre-training checkpoint, is free because someone published a weights file — 151 downloads at last count, no inference provider hosting it anywhere, and a model card that recommends the base "only for tasks that require heavy fine-tuning." One is a product with a meter on it. The other is raw material. This comparison is about which kind of free you actually need, because the two models will not work for the same person no matter how many months they both say "$0."

The names make the divergence easy to miss. Both sit in the "tiny / small model" category that has become the most crowded shelf in AI in 2026. Both are text-only, both carry contexts big enough for real work, and both will run a genuinely useful model on a modest budget. But they are built from different philosophies, owned by different parties, and rented or owned through completely different channels. Before you pick one, the single most useful question to ask is: do you need answers, or do you need weights?

Ling 3.0 Tiny, the free API

Ling 3.0 Tiny shipped on August 6, 2026, in a classic quiet-launch pattern: a commercial API listing from InclusionAI, Ant Group's model division, rather than a keynote. The architecture is a mixture-of-experts model with 7.9B total parameters and roughly 1.3B active per token — small enough to be served cheaply, big enough to hold real reasoning capability. It supports a 256K-token context window (the commercial listings show 262K), up to 32K output tokens, and it ships with two modes you can switch per request: a default-on "Thinking" mode that spends extra tokens reasoning, and an "Instant" mode that answers directly. Native function calling and prompt caching are both advertised, which matters for agent work.

The pricing is where the "free" needs unpacking. Today the model is genuinely free to call: it is listed as inclusionai/ling-3.0-tiny:free with a $0 rate on both input and output, and a second gateway is running it as a free promotion under inclusionai/ling-3.0-tiny-free until 8:00am PT on August 14, 2026. After that promotion, Vercel's listed rate card applies: $0.06 per million input tokens, $0.18 per million output tokens, and $0.01 per million cached-input tokens. That is a real price — roughly a third of the post-cut GPT-5.6 Luna tier, to give it scale — but it is not the "free" you can rely on past mid-August. Anyone building on the assumption that Ling 3.0 Tiny is permanently $0 is building on a promo that ends.

LFM2.5-2.6B-Base, the free weights

LFM2.5-2.6B-Base is the other kind of free entirely. It is the pre-trained base checkpoint behind Liquid AI's LFM2.5-2.6B family, published on Hugging Face on August 4, 2026 — the same week O​penAI cut GPT-5.6 Luna's price by 80%, which tells you something about why the whole market is crowded. The base is a 2.69-billion-parameter model in bfloat16 with a hybrid architecture: 22 double-gated short-convolution blocks plus 8 grouped-query attention layers, a 128,000-token vocabulary covering 16 languages, and roughly 34 trillion pre-training tokens — an unusually long training run for a model this size.

It is also, by every available measure, unmeasured. Liquid's own launch gave the post-trained sibling LFM2.5-2.6B the benchmark table and the 47,000 first-week downloads, while the base checkpoint got one parenthesis and 151 downloads. The base has zero published evaluations — no MMLU, no GPQA, no perplexity on any of Liquid's surfaces. It is text-only. Its context is 128K per the repository's config.json (the card advertises 131,072, and the config is what a training script actually packs to). And no inference provider hosts it, so the only way to see what these weights can do is to fine-tune and serve them yourself — which is exactly what the model card tells you to do. It ships under the LFM Open License v1.0, which caps commercial use at $10 million in annual revenue and carries that cap onto derivative works like your fine-tune.

The honest scoreboard

This matchup has the most lopsided evidence situation you will see between two 2026 models. Ling 3.0 Tiny has an independent score: Artificial Analysis gives it a 23 on the Intelligence Index, ranked #6 of 56 models in its size class, against a class median of 8 — well above average for a model this small. That index is a composite of nine evaluations, including Terminal-Bench 2.1, GPQA Diamond, and Humanity's Last Exam, run by a third party rather than selected by the vendor.

LFM2.5-2.6B-Base has no score at all. The agentic numbers people quote for the family — ToolSandbox 77.83, IFBench 59.17, Multi-IF 80.07, AIME25 51.87 — belong to the post-trained LFM2.5-2.6B, they are vendor-reported, and they are the output of a four-stage post-training pipeline that is not present in the base checkpoint. Take the base weights and you inherit the pre-training — the languages, the world knowledge, the long-context capability — and none of the instruction-following scoreboard. The single number the base carries is 34T training tokens, which is a budget, not a result. Nobody has yet reported what it bought.

The six dimensions that separate them, one line each:

• Price — Ling 3.0 Tiny $0.06 / $0.18 per 1M after the free promo (free until Aug 14) vs LFM2.5-2.6B-Base $0 to download, plus your training bill, forever.

• Context — 256K tokens on Ling 3.0 Tiny vs 128K on LFM2.5-2.6B-Base (config.json), a 2x gap.

• Parameters — ~1.3B active of 7.9B total MoE vs 2.69B dense; both text-only.

• Independent score — Ling 3.0 Tiny AA Intelligence Index 23 (#6/56) vs LFM2.5-2.6B-Base zero published evals.

• Modes — Ling 3.0 Tiny Thinking/Instant switch, function calling, prompt caching vs LFM2.5-2.6B-Base nothing until you post-train it.

• Served — Ling 3.0 Tiny live on its own commercial endpoints today vs LFM2.5-2.6B-Base hosted by no inference provider anywhere.

The asymmetry is the point: you can A/B Ling 3.0 Tiny against your own prompts this afternoon for pocket change, and the numbers you get back are checkable. The base checkpoint is a bet on your own fine-tuning skill with no external validation available yet.

The cost of "free": a metered answer vs a meter that stops

Ling 3.0 Tiny's meter never stops once the promo ends. Every token in and out bills at the listed rate, forever, at whatever concurrency you run. There is a hidden cost too: Artificial Analysis measured the model generating 210M output tokens while scoring its 56-model class — versus a 63M median — and flags it as "very verbose." A reasoning model that burns extra output tokens thinking has a higher effective cost per task than the rate card suggests, and the verbosity partially erodes the price advantage. Free-today pricing plus verbose-tomorrow metering is a real economic trade, not a rounding error.

The base checkpoint's "free" is a different accounting: it costs nothing to download, then costs you data, a training run, and hardware. A 100K-example supervised fine-tuning set is on the order of ten dollars in tokens to generate; a 1M-example set is closer to a hundred. The fine-tune itself — a 2.6B is a short run on a single GPU — is hours of machine time plus your engineering hours. To serve it you need a device; Liquid reports the family at roughly 220 tokens/second decode on an Apple M5 Max and 113 on a Ryzen AI Max+ 395, vendor-measured on the post-trained sibling, and the base has no speed numbers of its own. In practice that means a laptop you already own or a new machine starting around $1,500.

The decisive difference is the same one this blog used to frame LFM2.5-2.6B-Base against a repriced GPT-5.6 Luna: one meter never stops, the other stops the day training ends. Ling 3.0 Tiny charges per token at whatever your concurrency is, forever. The base's marginal token is free after the training run, at any concurrency, for as many concurrent agents as you can feed it. That is the only regime where a 2.6B truly beats a metered API on cost: high-volume, high-concurrency, always-on workloads.

Who should pick which

Pick Ling 3.0 Tiny if…

You need a working model this week: a capable text reasoning model you can call through an O​penAI-compatible endpoint, with a 256K context, tool calling, and an independent score you can look up. The free period gives you a real window to evaluate it against your own workload before paying anything, and even after the promo the $0.06 / $0.18 rate is aggressive for a model with a measured intelligence score above its class. You are renting intelligence with a meter, and you are comfortable with that.

Pick LFM2.5-2.6B-Base if…

You are building a model rather than renting one: a language or domain the off-the-shelf models under-serve, a regulated vertical where data cannot leave the device, post-training research, or a distillation student. Your workload is high-concurrency enough that zero marginal cost beats a metered bill, and you are comfortable under the license's revenue cap. You should also be honest with yourself about the risks: no published evals, no serving path, and a fine-tune you will have to debug without a benchmark to aim at.

The honest hybrid — and where a router fits

The most useful way to read this pairing is as two halves of one system rather than two rivals. The base checkpoint's real prerequisite is data: Liquid's own post-training recipe runs on teacher specialisation and on-policy distillation, which means making the base useful starts with a many-model data-generation job before it becomes a training job. That is exactly where a cheap, capable API model like Ling 3.0 Tiny — or, at higher quality, any of the text tiers on OrcaRouter — earns its place as the teacher for a small local student. Run the free window while it lasts, generate the fine-tuning corpus, then train the base into something you own.

On the API half, one honest caveat before we say anything about our own product: neither Ling 3.0 Tiny nor LFM2.5-2.6B-Base is on OrcaRouter. Ling 3.0 Tiny is served by its own commercial endpoints, and the base checkpoint is hosted by nobody. What OrcaRouter does usefully is carry the teacher models and the API traffic around them — Qwen 3.7 Flash at $0.03 / $0.13 per 1M, Gemini 3.5 Flash-Lite at $0.30 / $2.50, and 200+ more behind one key at 0% markup, so the provider's list price is what you pay and a vendor price cut lands the same day. For a data-generation run that must not die halfway, automatic failover across providers matters more than the model choice. But we do not host either of the two models this article is about, and we are not going to pretend otherwise.

Verdict

This is not a contest between two cheap models so much as a fork between two ways to spend your money. If you need answers — capable, checkable, rentable answers — Ling 3.0 Tiny is live today, free for the next week, independently scored above its class, and priced cheaply after the promo. If you want to own the model — offline, private, zero marginal cost, fine-tuned into something of your own — LFM2.5-2.6B-Base is one of the few genuinely unmeasured 2.6B checkpoints with 34T tokens of pre-training behind it, and you will be among the first people anywhere to find out how good it is. Both are free. They are free in different currencies: Ling 3.0 Tiny costs you a meter, and LFM2.5-2.6B-Base costs you a training run.