ONLINEAGENT_OPS 2026.Q3 HOME ARTICLES CRAFT RECORD BLOG MAP HUBS FAQ SEARCH
HOMEARTICLESGEMMA 4 VS MUSE GLIMMER: TWO OPEN MODELS, TWO
ARTICLES · COMPARISON

Gemma 4 vs Muse Glimmer: Two Open Models, Two Different Bets

Google shipped four sizes in April; Meta shipped one in August. Both Apache 2.0. What each was built for, and which questions neither vendor answers.

READ3 min
WORDS692
SECTIONS5
TYPECOMPARISON
CHECKED16 SEP 26

Same licence, different bet. Google shipped Gemma 4 as a family of four sizes on 2 April 2026 (five now, with a 12B added later); Meta shipped Muse Glimmer as a single 30B on 10 August 2026. Both are Apache 2.0. Google is covering phones to workstations; Meta is covering one job on one graphics card.

TL;DR — THE SHORT VERSION
  • Gemma 4 launched as four models — E2B, E4B, a 26B mixture of experts, and a 31B dense — and its model card now lists a fifth, a 12B.Google, Gemma 4 model card, read at source 22 Sep 2026: “The models are available in five distinct sizes: E2B, E4B, 12B, 26B A4B, and 31B.” First-hand: until 23 Sep 2026 this line said Gemma 4 “is four models”.Google, Gemma 4: Byte for byte, the most capable open models, 2 Apr 2026, read at source 16 Sep 2026: “We are releasing Gemma 4 in four versatile sizes”.
  • Muse Glimmer is one model, 30 billion parameters, aimed at local agents.Meta AI Research, Introducing Muse Glimmer, 10 Aug 2026, read at source 16 Sep 2026: “a 30-billion-parameter model optimized for always-on local agent workflows”.
  • Both are Apache 2.0, which is the part that changed for Meta — Llama never was.
  • Meta benchmarks itself against Gemma, not the other way round.
  • Neither claim is independent. Every number on this page is a vendor’s own.

Side by side

What each vendor published, read at source 16 Sep 2026
Gemma 4Muse Glimmer
Released2 Apr 202610 Aug 2026
SizesE2B, E4B, 26B MoE, 31B dense30B
LicenceApache 2.0Apache 2.0
Target hardwarephones to workstationsone consumer GPU
Built forreasoning and agentic workflowsalways-on local agent workflows
Images inyes, on the edge modelsyes, via a perception encoder

Google, Gemma 4, 2 Apr 2026 · Meta AI Research, Introducing Muse Glimmer, 10 Aug 2026. Both read at source 16 Sep 2026.

The difference that matters

Gemma 4 is a range. Google: “we’ve sized the Gemma 4 models specifically to run and fine-tune efficiently on hardware — from billions of Android devices worldwide, to laptop GPUs, all the way up to developer workstations and accelerators.” The small pair are built for phones: they “run completely offline with near-zero latency across edge devices like phones, Raspberry Pi, and NVIDIA Jetson Orin Nano”.

Muse Glimmer is a point, not a range. Meta: it is “a 30-billion-parameter model optimized for always-on local agent workflows”, and “small enough to run on a Mac or PC with a single consumer GPU”. The engineering behind that — 4-bit weights, a drafter model, a 24 GB target — is on how to run it locally.

A RANGE AND A POINT
Where each vendor says its model is meant to run. Positions are schematic, not measured.
Gemma 4 · five sizesE2B, E4BMuse Glimmer · one 30BONE CONSUMER GPUPHONESWORKSTATIONS
Reasoning — draws this page’s comparison of Google’s phones-to-workstations range and Meta’s single-GPU target, both from the vendor posts cited above. Page checked 16 Sep 2026.
TAKEAWAY

If the question is what you can put on a phone, Gemma has an answer and Glimmer does not. If it is what runs an agent on your desk all day, that is the thing Glimmer was sized for.

Why the licence is the news

Both models ship under Apache 2.0 — Google calls it “breakthrough capabilities made widely accessible under an Apache 2.0 license”, and Meta says it is “open sourcing the model weights under a permissive Apache 2.0 license”.

For Google that is continuity. For Meta it is a change of position: Llama shipped under Meta’s own community licence for years, and the state of that line is covered in is Llama dead.

What each says about the other

Meta names Google directly. Its own comparison is against “Gemma4-31B and Qwen3.6-27B”, and it claims Glimmer “performs strongly for its size class” against them. Google’s April post predates Glimmer and does not mention it.

Google’s own headline number is a leaderboard position: the 31B was “ranking as the #3 open model in the world on the industry-standard Arena AI text leaderboard” at the time it was published.

That was April. Leaderboard positions move monthly and this page has not re-queried it. Treat the rank as a dated snapshot, not a current fact.

What this page could not verify

  • Any benchmark. Both sets are self-reported and neither was re-run here.
  • The current leaderboard. Google’s rank is from 2 Apr 2026.
  • Real-world speed on any machine. Vendor figures only.
  • Whether the two are comparable at all. One is a family with a phone tier; the other is a single desktop-class agent model.
SOURCES

Google, Gemma 4: Byte for byte, the most capable open models, 2 April 2026 · Meta AI Research, Introducing Muse Glimmer, 10 August 2026. Both read at source on 16 September 2026.

Open-weight releases date fast.

ABOUTMETHODVERIFYPRIVACYCONTACTINDEXAI PROMPT GENEER · EVERY ARTICLE CARRIES ITS OWN CHECKED DATE