LFM2.5-24B-A2B vs Qwen3.8-27B: One Is a Download, the Other Is a Line Item

Start with the number that decides most of this. LFM2.5-24B-A2B is a file you put on a machine and keep. Qwen3.8-27B is a model you rent by the million tokens. Liquid AI announced the first on October 13, 2026 at 24.0 billion total parameters with 2.45B active per token and a 131,072-token context; A​libaba published the second on August 13, 2026 as a dense 27-billion-parameter vision-language model with a 262,144-token context. Both are open-weight. Only one of them has a published per-token price attached to a working endpoint, and it is not the smaller one.

That asymmetry is the comparison. Everything below — capability, licence, context, modality — hangs off whether you are buying hardware or buying tokens.

What each one is, and what it sees

LFM2.5-24B-A2B is a text-only mixture-of-experts: 40 layers split between 30 double-gated short-convolution blocks and 10 grouped-query attention layers, 64 experts per block with top-4 routing, a 128,000-token vocabulary, and 16 supported languages. Liquid's launch post counts pre-training at 33 trillion tokens; the Hugging Face model card for the same release says 36 trillion. The post-training pipeline ends in agentic reinforcement learning inside real harnesses, and the vendor's own guidance is to pair the model with retrieval or web search rather than asking it to answer knowledge questions from weights.

Qwen3.8-27B is a different shape entirely, and the shape is the point. It is a dense hybrid-attention design — 48 Gated DeltaNet linear-attention layers interleaved with 16 full-attention layers over 64 layers, hidden size 5,120, and a 248,320-token vocabulary — which is what makes a 262K context affordable in the first place. It takes text, image and video input and returns text. That mostly-linear backbone is inherited from the same lineage a compressed build like Ternary Bonsai 2 27B is derived from, which is a useful signal about how much of the value sits in the architecture rather than the training run.

The one-line version: Liquid's model routes and calls tools inside a loop, and does it without a network. Q​wen's model looks at things — screenshots, documents, video — and reasons about them at a context length roughly twice as long. Neither is a lesser version of the other; they are pointed at different halves of a working system.

The cost, worked as an actual number

This is where the two stop being comparable at all, so it is worth doing the arithmetic rather than trading adjectives.

On OrcaRouter, qwen/qwen3.8-27b is served at $0.33 per million input tokens and $2.40 per million output tokens, with a p50 time to first token of 3.45 seconds measured over seven days. A workload of 20 million input and 2 million output tokens a month — a modest internal tool — lands at $6.60 plus $4.80, or about $11.40. Nothing to install, nothing to keep warm, no GPU depreciation.

Liquid's Console endpoint runs $0.10 per million input with a $0.05 cached rate and $1.00 per million output, with $2 of starting credit, which prices the same workload at under $4. But that endpoint is not really the product. The product is the weights, and there is no vendor-published file-size table for LFM2.5-24B-A2B to quote — the quantization repos ship as siblings on Hugging Face. What the vendor does publish is the shape of the bet: a 2.45B-active MoE that runs at 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, both vendor-reported and unreproduced, and both meaningless on a datacenter GPU running many concurrent requests.

The break-even is not a token count, it is a utilisation question. Below roughly continuous heavy load, the metered endpoint wins on cash and loses on latency floor. Above it, the download wins on cash and loses on operations — you now own a deployment, its upgrades, and its failover story. That is the whole decision, and it has nothing to do with which model scores higher on a table.

Context, licence, and the two things that actually differ

• Context — 131,072 tokens versus 262,144 tokens. Q​wen's window is twice as long on paper, and the mostly-linear attention design is why it can afford it in memory. Liquid has published no long-context retrieval figures for its 131K window at any length, which is the one gap in its release that a benchmark round would close quickly.

• Licence — lfm1.0 versus Apache 2.0. Apache 2.0 has no revenue trigger and no field-of-use clause. Liquid's licence is its own instrument with its own terms. Read it against your company's numbers before a commercial rollout, and note that this difference does not shrink as the models get bigger.

• Modality — text only versus text, image and video in. If your workload involves a screenshot, a scanned page, or a video frame, this row ends the comparison by itself.

• Instruction following — IFEval 96.75 and IFBench 74.69. Liquid's strongest rows on its own launch table, ahead of every column in it. If the job is "obey this format exactly, every time," LFM2.5-24B-A2B has the better published case.

• Knowledge and coding — GPQA Diamond 74.42 and SWE-Bench Verified 51.00 for Liquid's model. Liquid compares those to Qwen3.6-35B-A3B's 83.69 and 66.20 and loses both. No Qwen3.8-27B comparison exists in that table at all, and nobody has run the two models against each other.

Where each one actually belongs

Choose Qwen3.8-27B when the input is not text, when the documents are long, or when you want to be billed rather than to operate something. A 262K window that takes images and video is a genuinely different product from a 131K window that takes text, and the Apache 2.0 licence removes a conversation most teams would otherwise have to have. The 3.45-second p50 time to first token is real and worth measuring against your own tolerance before you commit — a model this capable in the 27B dense class is not fast, and no architecture note changes that.

Choose LFM2.5-24B-A2B when the machine is the constraint. A 2.45B-active MoE with 131K context and an instruction-following profile that beats everything in its own comparison table is a strong thing to have running locally, and Liquid's decision to train it inside agent harnesses rather than against a static tool-calling set shows up in the rows it wins. What it is not is a document analyst or a knowledge base, and the vendor says so in the same document that publishes the benchmark table.

If you want to settle it on your own workload rather than on either vendor's table, the served leg is the cheap experiment: qwen/qwen3.8-27b is on our catalogue at the list price above with 0% markup on the provider's rate, and the sibling sizes qwen/qwen3.8-flash and qwen/qwen3.8-max sit alongside it behind the same key — so you can find the price-performance point your task actually needs before you decide whether it is worth owning a machine. LFM2.5-24B-A2B is not on our catalogue; Liquid's own Console and the open weights are the route to it, and nothing here should suggest otherwise.

The question that will settle this in six months is not which model wins. It is whether the local-executor tier holds up as context lengths grow — and on that, Q​wen's twice-as-long window and Liquid's unmeasured one are the only two data points anyone has.

Track live model outages and degradations in our incident archive and see which routes have failed over cleanly.