A hero title card reading 'LFM2.5-24B-A2B vs Gemma 4 26B-A4B' with the subtitle 'Two sparse models, two different budgets' and two labelled panels reading 'LFM2.5-24B-A2B: 24.0B total, 2.45B active, text only' and 'Gemma 4 26B A4B: 25.2B total, 3.8B active, text + image', with the OrcaRouter logo composited bottom-right.
Guides & Insights

LFM2.5-24B-A2B vs Gemma 4 26B-A4B: Two Sparse Models, Two Different Theories of Where the Compute Goes

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

LFM2.5-24B-A2B and Gemma 4 26B-A4B look like the same idea twice. Liquid AI's checkpoint shipped on October 13, 2026 at 24.0 billion total parameters with 2.45 billion read per token. G​oogle DeepMind's Gemma 4 26B-A4B carries 25.2 billion total with 3.8 billion active per token, published on Hugging Face on 11 March 2026. Both are mixtures of experts, both are roughly a 25-billion-parameter checkpoint attached to a two-to-four-billion-parameter compute budget, and both vendors reach for the same sentence when explaining the trade — you pay in memory for a big model and in latency for a small one. That near-identical arithmetic is exactly why the comparison is worth doing properly, because the two models then spend that budget on completely different things: Gemma 4 26B-A4B spends it on vision, a 256K context and a broad knowledge base; LFM2.5-24B-A2B spends it on a reasoning trace, a tool-calling harness and agentic reinforcement learning. Same shape, opposite priorities.

There is one more asymmetry sitting on top: the LFM2.5-24B-A2B benchmark table uses Gemma-4-26B-A4B-it as one of its three baselines, so for once we are not lining up two launch tables and guessing — Liquid ran both, in the same vLLM environment, and published the rows.

Both are missing the same thing, and it is the thing everyone assumes

Start with the misconception these two models are most likely to generate, because it is the same one for both. A sparse mixture of experts does not shrink your memory requirement. Every parameter in the checkpoint has to be resident; what 2.45B or 3.8B active buys you is throughput per token, not a smaller footprint. A reader who sees "2.45B active" and reaches for a 4 GB machine is going to have a bad afternoon.

On our own catalogue the Gemma 4 26B A4B checkpoint is listed with a 262,144-token context, and the LFM2.5-24B-A2B card gives 131,072. On disk the gap is bigger than the parameter counts suggest: LFM2.5-24B-A2B publishes a Q4_K_M GGUF at 14.5 GB and a QAT QAD-Q4_0 at 13.6 GB, with 48 GB in BF16; Gemma 4 26B A4B publishes bfloat16 safetensors as a 49.9 GB shard plus a 1.7 GB shard, with a quantisation-aware 4-bit GGUF pairing a compact weights file with a separate multimodal projection file. Round numbers: about 14 GB against about 50 GB uncompressed, from two checkpoints whose total parameter counts differ by one billion. That difference comes from the vocabulary, the expert layout and the multimodal stack, not from the parameter headline, and it is the single most practical line in this whole piece.

The spec contrast, one line per dimension

• Total / active parameters — LFM2.5-24B-A2B 24.0B / 2.45B, 64 experts with top-4 routing across 40 layers. Gemma 4 26B-A4B 25.2B / 3.8B, 8 active experts out of 128 plus one shared expert across 30 layers.

A capture of the Hugging Face model page for google/gemma-4-26B-A4B-it, showing the image-text-to-text pipeline tag, the Transformers and Safetensors formats, the gemma4 and conversational tags, the Apache 2.0 licence, and the card text stating that Gemma 4 models are multimodal, handling text and image input with audio supported on the E2B, E4B and 12B models, that the family spans a context window of up to 256K tokens across five sizes, and the architecture list covering reasoning, extended multimodalities, dense and MoE variants, and on-device optimisation.

• Context — 131,072 tokens versus 262,144, with Gemma 4 26B A4B using a 1,024-token sliding window and hybrid attention that interleaves local and global layers, keeping the last layer global.

• Modality — LFM2.5-24B-A2B is text in, text out. Gemma 4 26B A4B takes text and image, with a vision encoder of roughly 550M parameters; audio is confined to the smaller Gemma 4 sizes.

• Reasoning — LFM2.5-24B-A2B always emits an explicit chain of thought inside a <think> block. Gemma 4 26B A4B is an instruction-tuned generalist with no mandatory trace.

• Vocabulary — 128,000 tokens, doubled from the previous generation by in-place tokenizer expansion, versus a 262K family vocabulary shared across all five Gemma 4 sizes.

• Licence — lfm1.0, free for commercial use until $10M in annual revenue and a commercial licence past that. Gemma 4 26B A4B is Apache 2.0, the stock licence text at ai.google.dev/gemma/apache_2, with G​oogle's prohibited-use policy sitting alongside it as an acceptable-use document rather than a licence restriction.

• Published formats — GGUF (Q4_K_M through BF16), FP8, ONNX and 4/5/6/8-bit MLX for the Liquid model; bfloat16 safetensors, a pretrained checkpoint, a QAT 4-bit build and a llama.cpp-ready GGUF, plus a multi-token-prediction drafter, for the Gemma size.

Liquid ran both, and the results are not a clean sweep

The LFM2.5-24B-A2B launch post puts Gemma-4-26B-A4B-it in the baseline column of its sixteen-benchmark table, alongside Qwen3.6-35B-A3B and Nemotron-3.5-Lightning, with all four columns run in vLLM under per-benchmark generation settings. Everything below is vendor-reported on both sides — Liquid's numbers, in a table Liquid assembled. It is not an independent evaluation, but it is at least a single harness, which is more than most model-versus-model pages get.

The agentic rows read like a rout: AutomationBench 27.69 against 5.46, ToolSandbox 80.71 against 72.09, LiveClaw 48.09 against 41.25, τ³-Bench Banking 8.04 against 4.54, BFCLv4 65.01 against 57.09. Instruction following is the same story — IFEval 96.75 against 92.00, IFBench 74.69 against 47.44, Multi-IF 84.38 against 82.19.

Then the other half of the table, which Liquid published anyway: PinchBench 83.18 against 77.12 goes the other way, and so do three rows that matter more than any of the tool-calling wins. GPQA Diamond — 74.42 against 74.90, essentially level at the top end of graduate science. SWE-Bench Verified via OpenCode — 51.00 against 25.80, a two-to-one margin in the Liquid model's favour, which is the most surprising single number in either table. SciCode — 32.71 against 35.42, the Gemma size ahead. And AA-Omniscience — −24.90 against −59.39, where both models score badly and the Gemma size scores more than twice as badly.

Two readings of that, and the honest one is the second. The first is that a 2.45B-active agent model beats a 3.8B-active generalist at agents, which is what it was built to do. The second is that the Gemma 4 26B A4B column looks anomalously weak across the board — 25.2B total parameters and 3.8B active landing under a 2.45B-active model on instruction following, tool use and knowledge is not what G​oogle's own family table would predict, and it is a reminder that a shared harness does not make a benchmark comparable if the models are driven through different generation settings and prompting. Treat the direction of these rows as informative and the magnitudes as not.

What is not in the table at all is the multimodal half of Gemma 4 26B A4B. A vision benchmark is impossible here because LFM2.5-24B-A2B cannot see. If any part of your workload involves an image arriving from a user, this comparison has a winner before it starts, and it is not the newer model.

Price, and what each one costs to actually hold

LFM2.5-24B-A2B has a per-token price: $0.10 per million input tokens ($0.05 cached) and $1.00 per million output on Liquid's own Console, with $2 of credit to start. Gemma 4 26B A4B has no vendor list price at all — G​oogle publishes none — so the only honest cost figure is whatever the platforms hosting the open weights charge, which varies by host and by day.

That asymmetry is the argument for going through a router rather than wiring one vendor's endpoint into your code. Geometry aside, this is the point where a checkout for both sides matters: we serve the Gemma 4 26B A4B checkpoint under two ids — google/gemma-4-26b-a4b-it and obsidian/gemma-4-26B-A4B — and both price references Liquid uses in its table, qwen/qwen3.6-35b-a3b among them, behind one key at 0% markup on the provider's list price. We do not serve LFM2.5-24B-A2B. The practical version of this page is therefore: pull the Liquid checkpoint down and run it locally, and call the Gemma 4 size and the Q​wen baseline through the same key while you are deciding, so you are comparing on your own tasks rather than on either vendor's choice of harness.

A two-column scoreboard titled 'LFM2.5-24B-A2B vs Gemma 4 26B A4B — the scoreboard'. The left column, LFM2.5-24B-A2B, lists 24.0B total and 2.45B active parameters, a 131,072-token context, text-only input, SWE-Bench Verified 51.00, AutomationBench 27.69, and a Q4_K_M download of 14.5 GB. The right column, Gemma 4 26B A4B, lists 25.2B total and 3.8B active parameters, a 262,144-token context, text and image input, SWE-Bench Verified 25.80, AutomationBench 5.46, and bfloat16 weights of about 50 GB, above a footer reading 'Benchmark rows from Liquid AI's 2026-10-13 vLLM table, which ran both models; vendor-reported and unreproduced.' and the OrcaRouter logo composited bottom-right.A capture of OrcaRouter's own model page for google/gemma-4-26b-a4b-it, showing the 'Gemma 4 26B A4B' heading and the google/gemma-4-26b-a4b-it identifier, the vision, tools, JSON and reasoning chips, Google as the vendor with a 2026-04-03 date, the summary noting 25.2B total parameters with only 3.8B activating per token, and the 262K-token context row listing text, image and video input with a p50 time-to-first-token figure.

Which one, and when the question is really about your machine

If your workload is text and your agent needs to plan, call tools and explain itself, the newer checkpoint is the one to download: 14.5 GB at Q4_K_M, a 131K context, explicit reasoning, and a benchmark table where it wins the agentic rows against this exact opponent. If any of your work involves images, or you need 256K of context rather than 128K, or you want an Apache 2.0 licence with no revenue clause to track, Gemma 4 26B A4B answers a question LFM2.5-24B-A2B does not even ask — at roughly three and a half times the uncompressed disk footprint.

The thing neither model fixes is that both are memory-bound rather than compute-bound, and both vendors have to be read the same way. G​oogle's family launch post describes the two larger Gemma 4 sizes as bfloat16 weights that "fit efficiently on a single 80GB NVIDIA H100 GPU" — true, and also the vendor's own framing rather than a measurement of this specific build. Liquid's 184 tokens per second on an Apple M5 Max and "up to 1.7× faster than comparable MoEs on-device" carry no reproduction either. Two sparse models, two vendors, one shared instruction: the arithmetic in the name tells you about speed, and nothing in the name tells you about memory.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.