LBS-27B vs Gemma 4 26B A4B: What an Empty Repository Has to Beat

If your workload involves an image, a scanned page or a screenshot, the comparison between LBS-27B and Gemma 4 26B A4B is already over. Gemma 4 26B A4B accepts images and video frames and returns text; LBS-27B does not yet exist in any runnable form, so it accepts nothing. Everything else in this matchup is a question about what inclusionAI's empty repository might become, and the honest way to answer it is to describe the bar Gemma has already set rather than pretend the two models have been measured against each other.

That asymmetry is the whole story. One of these models has been on Hugging Face since March 2026, has 8.1 million downloads, a full evaluation table and a published price. The other was created at 05:36 UTC on 26 October 2026 and contains a licence header and nothing else. A side-by-side that treats them as peers would be fiction. A side-by-side that tells you what LBS-27B would have to beat is useful, and that is what follows.

The known side of the ledger

Gemma 4 26B A4B is a mixture-of-experts model with 25.2 billion total parameters and 3.8 billion active per token, drawn from 128 experts with eight active plus one shared. It has 30 layers, a vocabulary of 262,144 tokens and a 256K context window. The attention stack alternates five sliding-window layers of 1,024 tokens with one full-attention layer, which is how Google keeps the long window affordable.

It is multimodal in a specific, limited way. There is a vision encoder of roughly 550 million parameters, image and video-frame input, and text output. Audio is not supported on this model — it lives on the smaller E2B, E4B and 12B variants. The card claims support for more than 140 languages with 35 or so working out of the box. The licence is Apache-2.0, and the training cutoff is January 2025.

The benchmark table on Google's model card — vendor-reported, not independently reproduced — reads as follows: MMLU Pro 82.6, GPQA Diamond 82.3, AIME 2026 without tools 88.3, LiveCodeBench v6 77.1, Codeforces ELO 1718, tau2 averaged over three suites 68.2, MMMU Pro 73.8, MATH-Vision 82.4. The weak rows are as informative as the strong ones: Humanity's Last Exam scores 8.7 without tools and 17.2 with search, and MRCR v2 at 128K with eight needles averages 44.1. A 256K window does not mean the model retrieves well across 256K.

The unknown side, stated plainly

LBS-27B is a repository, a licence and a name. The repository was created on 26 October 2026 and contains two files: a standard .gitattributes and a 28-byte README whose entire content is the line license: apache-2.0. There is no config.json, no weight shard, no tokenizer and no model card. There are no recorded downloads and zero likes. We could find no announcement, paper or social post anywhere.

What that means for this comparison is simple: every row other than the licence is blank. The "27B" is a string in the repository name, not a verified parameter count. The architecture could be dense or sparse. The context length could be anything. There may or may not be a vision encoder. Until a config file appears, none of those questions has an answer, and any article that fills them in is inventing.

Where the matchup would actually be decided

If LBS-27B ships as a 27B-class model, these are the dimensions that will separate it from Gemma 4 26B A4B. Each line carries both sides, with the unknown marked as unknown rather than left blank.

• Total and active parameters — Gemma 4 26B A4B is 25.2B total and 3.8B active, vendor-reported; LBS-27B is an unverified 27B with no active-parameter figure.

• Architecture — Gemma uses 128 experts with top-8 routing; LBS is unknown, and dense versus sparse is the single most consequential unknown for serving cost.

• Context window — 262,144 tokens for Gemma; unknown for LBS.

• Modality — text, image and video-frame input with text output for Gemma; unknown for LBS, and text-only is the base case until a vision tower appears.

• Licence — Apache-2.0 for Gemma; Apache-2.0 declared for LBS, though a licence can change before weights ship.

• Price — $0.06 per million input tokens and $0.33 per million output tokens for Gemma on OrcaRouter, at 262,144 tokens of context; no price exists for LBS because there is nothing to serve.

• Published evaluations — a full vendor table for Gemma; none for LBS.

The number that decides the economics: 3.8 billion

The A4B in Gemma's name is the most important part of it. At roughly 3.8 billion active parameters per token, the model does the compute of a small model while holding the knowledge of a 25B one. That is why it can be priced at six cents per million input tokens and still be served commercially, and it is why the model runs on hardware that a dense 25B would struggle with.

A dense 27B is a different product. It reads every parameter for every token, so its decode cost scales with the full weight count. If LBS-27B turns out to be dense, it will not compete with Gemma 4 26B A4B on price at equal quality — it will compete on quality at a higher price, which is a much harder argument in 2026. If it turns out to be sparse with a small active count, the comparison becomes real and interesting.

This is why the first credible artifact to watch is not a benchmark score. It is the expert count in the config file. Everything about the commercial shape of LBS-27B follows from that one field.

Vision is a workload class, not a feature

It is easy to treat Gemma's image input as a checkbox. In practice it changes which jobs the model can hold at all. Document extraction, screenshot triage, chart reading, UI regression checks, invoice processing and OCR-adjacent pipelines are not text tasks with a picture attached; they are tasks where the picture is the input. A text-only model cannot enter them at any benchmark score.

Gemma's vision numbers are respectable rather than dominant: MMMU Pro 73.8, MATH-Vision 82.4 and an OmniDocBench 1.5 average edit distance of 0.149, all from Google's card. The visual token budget is selectable at 70, 140, 280, 560 or 1,120 tokens per image, which lets you trade resolution against cost per request — a practical detail that matters more in production than a leaderboard point.

For LBS-27B, the question is binary. If inclusionAI ships a vision tower, the model enters the same workload class and the comparison becomes a contest of quality and price. If it does not, LBS-27B can only take the text half of Gemma's job, and the decision becomes a routing question rather than a replacement question.

What LBS-27B would have to clear

Assume, for the sake of argument, that the weights arrive and the model is a competent 27B. To be worth switching to from Gemma 4 26B A4B on text workloads, it would need to do three things at once.

First, beat Gemma's vendor-reported rows by a margin that survives harness differences. MMLU Pro 82.6, GPQA Diamond 82.3 and AIME 2026 at 88.3 are strong numbers for a model with 3.8B active parameters, and a two-point lead on a different harness version is not a lead. Second, do it at a comparable blended price. Gemma's output price of $0.33 per million is the binding constraint; a dense 27B at three to five times that output rate needs a much larger quality gap to justify itself. Third, beat it on something Gemma is visibly weak at — long-context retrieval, where Gemma's 44.1 on MRCR v2 at 128K leaves real room, or tool use, where its tau2 average of 68.2 is good but not exceptional.

The most plausible way for a new 27B to win is not to beat Gemma everywhere. It is to be clearly better at one expensive job — long-document analysis, agentic tool chains, or a non-English language where Gemma's coverage is thin — while being merely competitive elsewhere.

Price, availability and the practical answer

Gemma 4 26B A4B is available on OrcaRouter today as google/gemma-4-26b-a4b-it, with a 262,144-token context and pricing at $0.06 per million input tokens and $0.33 per million output tokens. Because OrcaRouter passes through provider list price at 0% markup, any price change Google or its hosting partners make shows up on our side the same day rather than at the next billing cycle. There is also a second Gemma 4 26B A4B listing on the platform at a higher rate, which is worth knowing before you assume the cheapest line is the only one.

LBS-27B is not on OrcaRouter, and we will not pretend it is. There are no weights to serve. When and if that changes, the sensible way to evaluate it is not a wholesale migration but a split: route a slice of traffic to the new model through the same API key, keep Gemma as the fallback, and compare on your own inputs. Automatic failover makes that safe in a way a direct integration does not, because a bad or unavailable model degrades to a known one instead of taking the request down.

Who should act, and who should wait

If you need a cheap multimodal model with a long context and a permissive licence this week, Gemma 4 26B A4B is the answer and LBS-27B is not in the conversation. That is not a knock on inclusionAI; it is what an empty repository means.

If you are watching the 27B class for a text-only workhorse and you can afford to wait, the thing to watch is the config file. A sparse architecture with a small active count would make LBS-27B a genuine price competitor to Gemma. A dense one would make it a quality play at a higher cost. And a vision tower in the repository would change the comparison from "text alternative" to "direct rival" overnight.

Until one of those appears, the most useful thing this page can say is the least exciting: Gemma has a measured spec sheet and a live price, LBS has a licence line, and the gap between them is not a gap in quality but a gap in evidence.