MiniCPM5-2B-DSpark vs LFM2.5-2.6B-Base: One Is an Ingredient for Serving, One Is an Ingredient for Training

Two of the most interesting small-model releases of late summer 2026 have this in common: neither works out of the box, and both vendors say so in the first sentences of their cards. MiniCPM5-2B-DSpark is the DSpark draft checkpoint OpenBMB posted quietly to Hugging Face on September 6, 2026 — a 323.8M-parameter companion to MiniCPM5-2B, the dense 2.52B on-device model ModelBest announced at WAIC on July 19, 2026 — and it does nothing until you load it next to that target in a speculative-decoding engine. LFM2.5-2.6B-Base is Liquid AI's pre-trained text checkpoint, on Hugging Face since August 1, 2026, which Liquid's own card calls "the pre-trained text-only checkpoint" recommended only for tasks requiring heavy fine-tuning. One is an ingredient for serving a finished model faster; the other is an ingredient for training your own model in the first place. They sit on different shelves, and the builder's mistake would be reaching for the wrong one.

What makes them worth comparing is that they target the same hardware and the same 2026 ambition — capable AI on phones, laptops, and edge boxes — from opposite ends of the build process. If you want a deployable on-device model plus a free speed-up, OpenBMB's stack hands you the whole thing open. If you want to own the post-training yourself, Liquid hands you the substrate and the labor. This article is about which layer you are actually shopping for.

Two cards that say "not for direct use" in different dialects

Read the two model cards side by side and the tell is in the pairing language. MiniCPM5-2B-DSpark's card is written entirely in terms of another model: a target, a tokenizer it shares, draft layers, acceptance lengths, an SGLang launch command that takes a target path and a draft path. The draft has no identity of its own — five full-attention layers, 323,776,001 parameters, a block size of seven, trained for six epochs on 7.05 billion tokens of MiniCPM5-2B-generated responses, and released under Apache-2.0 with the confidence and Markov heads of the DSpark method (arXiv 2607.05147) enabled. Its purpose is throughput: propose up to seven tokens, have the 2.52B target verify the block in one pass, keep what it accepts.

The openbmb/MiniCPM5-2B-DSpark card above is the full public surface of the release — a serving accessory with no announcement attached, reporting acceptance length but no end-to-end speedup.

LFM2.5-2.6B-Base's card is written in terms of a model that does not exist yet: yours. The 2.69B hybrid stacks eight grouped-query-attention layers on twenty-two dual-gated short-convolution blocks — Liquid's architecture for on-device deployment — trained on roughly 34 trillion tokens across 16 languages with a 128K context. It is text-only, has no instruction tuning, and publishes zero benchmarks of its own, because a pre-trained base is not scored; the fine-tuned products that come out of it are. Its job is to be continued-pretrained or SFT'd on your data, after which anything you ship inherits its license terms.

The LiquidAI/LFM2.5-2.6B-Base card above shows the family table — base versus the post-trained LFM2.5-2.6B agentic model — and the "pre-trained text-only checkpoint" note that defines the base as raw material.

The repos, dimension by dimension

• Role — MiniCPM5-2B-DSpark: accelerates an existing, finished target. LFM2.5-2.6B-Base: the substrate you post-train into a finished model.

• Parameters — MiniCPM5-2B-DSpark: 324M draft for a 2.52B dense target. LFM2.5-2.6B-Base: 2.69B hybrid (attention + short-convolution blocks).

• Context — MiniCPM5-2B target: 128K native. LFM2.5-2.6B-Base: 128K native.

• What you must add — MiniCPM5-2B-DSpark: a serving engine (SGLang DSPARK) and the target model. LFM2.5-2.6B-Base: a fine-tuning run, an eval suite, and an inference stack.

• Benchmarks — MiniCPM5-2B-DSpark: repo acceptance length only (greedy aggregate 5.52 per step), no speedup figure. LFM2.5-2.6B-Base: none published for the base itself.

• License — MiniCPM5-2B-DSpark: Apache-2.0, no strings. LFM2.5-2.6B-Base: LFM Open License v1.0 — free commercial use up to $10M annual revenue, and the cap is inherited by whatever you fine-tune.

The scoreboard above is the ingredient shelf drawn as a card: one column is a serving add-on for a finished model, the other a training substrate for a model that does not exist until you build it.

The $10 million difference

For a builder, the licensing gap is the least glamorous and most decisive row on that list. ModelBest and OpenBMB put MiniCPM5-2B and its draft under Apache-2.0: you can serve the target, run the draft, and build a commercial product on top with no revenue ceiling and no inherited obligation beyond attribution. Liquid's LFM Open License v1.0 permits commercial use but caps it — free below $10M in annual revenue, and the cap travels into your derivative work, so a product you fine-tune from LFM2.5-2.6B-Base carries the same ceiling. That does not make Liquid's license unreasonable; it makes it a term a startup hoping to scale past the cap has to plan around from day one. If your project's revenue ambition is small or you are evaluating research, the cap is a non-issue. If you are building something you intend to grow, Apache-2.0 is a different kind of raw material.

The build-versus-buy fork for builders

Frame it as a build-versus-buy decision about the model itself, and the choice stops being confusing. MiniCPM5-2B-DSpark is the "buy" path: the hard work — post-training an on-device model for agentic, long-context work — is already done in MiniCPM5-2B, and the draft is a no-extra-license acceleration layer on top. You are buying a finished model and a speed-up, both Apache-2.0, and your remaining work is serving and evaluation. The open question is quantitative: OpenBMB has measured the draft's acceptance (5.52 tokens per verification step at greedy) but not its wall-clock payoff, so you still have to benchmark the SGLang deployment on your own hardware to know what you bought.

LFM2.5-2.6B-Base is the "build" path: you are buying the pretraining substrate and doing the post-training yourself, which is the only honest reason to choose a base over the post-trained LFM2.5-2.6B sibling that Liquid already sells. That path makes sense when you have a domain, a language, or a behavior profile that off-the-shelf models do not cover — you take Liquid's architecture and 34-trillion-token foundation and make it yours. It is strictly more work, it starts with zero published benchmarks, and it carries the revenue cap forward. The payoff is a model no one else has.

The routing reality

Neither checkpoint appears in any hosted catalog today, OrcaRouter's included — the draft is a local serving accessory by definition, and Liquid's base is a self-host fine-tuning project. That is where a routing layer stays useful without pretending to host either one: once you have served or fine-tuned either model, the layer in front of your own endpoints is what makes trialing them reversible. Point a test path at a MiniCPM5-2B-plus-draft SGLang build while production stays on whatever you serve today, let automatic failover catch the stall, and keep provider list prices passing through at 0% markup for the hosted models you compare against. The point of the layer here is not delivery — it is that neither ingredient commits you irreversibly.

Which shelf are you shopping on

Choose MiniCPM5-2B-DSpark (with its target) when you want a finished, Apache-2.0 on-device model today and a draft that may make it faster — accepting that you must measure the draft's speedup yourself, since the vendor has not. Choose LFM2.5-2.6B-Base when you intend to post-train your own model and want Liquid's hybrid architecture and 16-language foundation as the starting weights — accepting the fine-tuning bill, the zero published benchmarks, and the $10M revenue cap that will ride along in whatever you ship. The two are not rivals for the same job; they are two different jobs on the same hardware, and the only wrong choice is buying one while believing you were buying the other.