5.0× less energy than Jamba 1.5 Large.
Jamba 1.5 Large Environmental Impact
Hybrid SSM-Transformer architecture for efficient long context
One query, in context
Energy, carbon and water for a single use. Short query (~300 tokens)
About 1.2× a minute of laptop use and 8% of a phone charge.
About 3.8× a Google search and 21% of 10 m by car.
About 42× a drop of water and 42% of a teaspoon of water.
- Where it runsCloud (US)
- Grid carbon intensity450 g CO₂/kWh · 25% renewable
- Compared with a Google search3x more than a search (0.3 Wh)
- Water~476 queries to fill 1 litre
Where Jamba 1.5 Large sits
#139 of 199 by energy per query. Heavier than 69% of all models. Lit dots are the same kind of model; click any dot to open it.
Want something lighter?
1.6× less energy than Jamba 1.5 Large.
2.5× less energy than Jamba 1.5 Large.
What a month of use adds up to
Pick how much you use it. Based on one quick use at a time, five days a week.
- Energy
- 108.3 Wh
- Carbon
- 48.7 g
- Water
- 227 mL
A month of 5 chats a day, 5 days a week. The energy is about 9.0× a phone charge and 98% of a kettle boil.
Work out your own →- Energy
- 433.0 Wh
- Carbon
- 194.8 g
- Water
- 909 mL
A month of 20 chats a day, 5 days a week. The energy is about 3.9× a kettle boil and 6% of a UK home's day.
Work out your own →- Energy
- 2.17 kWh
- Carbon
- 974.3 g
- Water
- 4.55 L
A month of 100 chats a day, 5 days a week. The energy is about 20× a kettle boil and 29% of a UK home's day.
Work out your own →How does Jamba 1.5 Large compare?
Next to similar models. Bars use a log scale, so each step is a big jump.
Detailed breakdown
Energy
Jamba 1.5 Large uses a novel SSM-Transformer hybrid architecture that combines Mamba state-space layers with traditional attention. This gives it significantly better efficiency for long contexts — O(n) rather than O(n²) scaling. At ~1.0 Wh for short queries, it's competitive with smaller models, and the efficiency advantage grows with longer contexts.
Power source & carbon
Available as open-source or via AI21 API. The hybrid architecture makes it particularly attractive for long-context inference on more modest hardware.
Water use
At ~2.1 mL per query, Jamba is water-efficient relative to its 398B parameter count, thanks to the SSM components reducing overall compute.
About Jamba 1.5 Large
Jamba 1.5 Large is a 398B-parameter text and chat model from AI21 Labs, released August 22, 2024. Hybrid SSM-Transformer architecture for efficient long context. At 1.0 Wh per query, it uses 3x the energy of a Google search. It runs on a SSM-Transformer Hybrid (Mamba + Attention) architecture.
These figures are estimates derived from hardware specifications and API benchmarks — AI21 Labs has not published official energy data for Jamba 1.5 Large. Actual consumption may vary significantly depending on batching, quantisation, and infrastructure optimisations that we cannot observe from outside.
Jamba 1.5 Large in context
Key insights
- Open-source weights — can be self-hosted on infrastructure you control
What does your Jamba 1.5 Large usage cost the planet?
Use the calculator to estimate your personal footprint based on how often you use Jamba 1.5 Large.
Frequently asked questions
How much energy does Jamba 1.5 Large use per query?
Each Jamba 1.5 Large query consumes approximately 1.0 Wh of energy. This is 3x more than a traditional Google search (~0.3 Wh).
What is Jamba 1.5 Large's carbon footprint?
Based on the carbon intensity of Cloud (US), each query produces approximately 0.45 g of CO2. The grid in this region has a carbon intensity of 450 g CO2/kWh with 25% renewable energy.
How much water does Jamba 1.5 Large use?
Each query consumes approximately 2 mL of water, primarily used for cooling the data centers that process the request.
How does Jamba 1.5 Large compare to a Google search?
A Jamba 1.5 Large query uses 3x more than a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while Jamba 1.5 Large uses 1.0 Wh.
Technical details
- Architecture
- SSM-Transformer Hybrid (Mamba + Attention)
- Parameters
- 398B
- Context window
- 256,000 tokens
- Release date
- 2024-08-22
- Open source
- Yes
- Training data cutoff
- 2024-08