Software decision profile

Gemma 3 (27B)

Google's 27B open-weights model, tuned to fit machines with 32GB of unified memory.

AI Models & IntelligenceGemma 3 (27B)Decision profile Fathom Layer

Evidence statusSpecs not independently verifiedhuggingface.co

Visit site

Editorial assessment paid

Explore the trade-offs.

Where it stands outFits comfortably in 32GB of unified memoryExplore strengths ↗Close details ↑

Matches or beats the previous generation of 70B models
Native multimodal capability at this size

What to weighGoogle's licence carries commercial restrictions worth legal reviewExplore limitations ↗Close details ↑

Still needs mid-tier hardware at minimum
Weaker fine-tuning ecosystem than Llama

Best suited toStartups needing capability on modest hardware · Local researchers on consumer machines · Multimodal work without a server

Where it wins, where it doesn't

Pros

Fits comfortably in 32GB of unified memory
Matches or beats the previous generation of 70B models
Native multimodal capability at this size

Cons

Google's licence carries commercial restrictions worth legal review
Still needs mid-tier hardware at minimum
Weaker fine-tuning ecosystem than Llama

Ideal forStartups needing capability on modest hardwareLocal researchers on consumer machinesMultimodal work without a server

Editorial note Read note ↓

Gemma 3 is the clearest evidence that parameter count stopped being a proxy for capability. A well-optimised 27B model now matches or beats the previous generation of 70B models on most tasks, and — the part that actually changes what you can do — it fits comfortably on a 32GB unified-memory machine. That moves serious local inference from a workstation purchase to a laptop you might already own. Native multimodality is included rather than bolted on, which is still uncommon at this size. The thing to check before committing is the licence: Google's terms carry restrictions that a genuinely permissive licence does not, and for a commercial product that is a legal question, not a technical one. Read it before you build on it.

Editorial analysis

Capabilities, limitations and fit.

In-Depth Review

Why that matters

It moves serious local inference from a workstation purchase to a laptop you might already own. Native multimodality is included rather than bolted on, which is still uncommon at this size.

The thing to check first

The licence. Google's terms carry commercial restrictions that a genuinely permissive licence does not. For a commercial product that is a legal question, not a technical one — read it before you build on it. The fine-tuning ecosystem is also thinner than Llama's.

Who should use it

Startups needing real capability on modest hardware, local researchers on consumer machines, and multimodal work without a server. If you need a fully permissive licence or the deepest tooling, Llama is the safer base.

Frequently Asked Questions

Who is Gemma 3 (27B) for?↓

Gemma 3 (27B) is a fit for startups needing capability on modest hardware, Local researchers on consumer machines and Multimodal work without a server.

What are the drawbacks of Gemma 3 (27B)?↓

The trade-offs we record are: Google's licence carries commercial restrictions worth legal review, Still needs mid-tier hardware at minimum and Weaker fine-tuning ecosystem than Llama.

What does Gemma 3 (27B) do well?↓

Fits comfortably in 32GB of unified memory, Matches or beats the previous generation of 70B models and Native multimodal capability at this size.

Alternatives to consider

See all alternatives →

Continue the inquiry