source:- www.bing.com
Introducing Gemini 2.0 Flashlight
Google has expanded the AI Gemini 2.0 Flash family with the addition of Flashlight, a new model, and a new Pro version.
They’ve also released a new experimental feature, Flash Thinking, which can reason across YouTube, Maps, and Search. Access to YouTube and Maps is something Open AI’s models currently lack.
The Appeal of Flashlight
The biggest release of the three is the Pro version. However, I’m personally more interested in AI Gemini 2.0 Flashlight, though limited information is available. It’s described as a new, cost-efficient variant.
The cost is approximately 7 cents per million input tokens and about 30 cents per million output tokens, which is remarkable.
We’ll directly compare the AI Gemini 2.0 Flash and AI Gemini 2.0 Flashlight models later, but first, let’s see how all three stack up against each other.
source:- www.bing.com
Gemini 2.0 Model Comparison
All three are multimodal for input, processing audio, images, and text. Currently, output is limited to text, but image and audio output for Gemini 2.0 Flash and Pro are coming soon.
The Light version doesn’t appear to have this capability. Regarding context window, Pro has the longest, at 2 million tokens, compared to 1 million for AI Gemini 2.0 Flash and AI Gemini 2.0 Flashlight.
All three support structured output and function calling, although I don’t see function calling with the Gemini 2.0 Flashlight preview in AI Studio.
Gemini 2.0 Flashlight will not have search as a tool or code execution behind the API.
Flashlight’s Potential
AI Gemini 2.0 Flashlight seems significantly improved compared to the previous 1.5 AI Gemini 2.0 Flash, even surpassing the original 1.5 Pro in some cases.
The combination of cost-effectiveness and performance makes it a very interesting choice, potentially usable as a workhorse model.
Often, the most powerful model isn’t necessary; tasks require a reasonably good model, and Gemini 2.0 Flashlight seems to fit that bill.
source:- www.bing.com and Canva
But how does it compare to AI Gemini 2.0 Flash? In this blog post, we’ll test it on a few programming and retrieval tasks. To compare these models, I’ll use the compare option in AI Studio.
The first model will be AI Gemini 2.0 Flash, and the second, Gemini 2.0 Flashlight preview. First, I’ll test long context. I’ve uploaded the DeepSeek-3 technical report, which is 53 pages long. Before testing, I’ll set the temperature to zero for both.
Testing and Results
Looking at the available tools, Gemini 2.0 Flash has access to structured output, code execution, function calling, and Google Search grounding.
AI Gemini 2.0 Flashlight only has access to structured output. My prompt is: “What are the major contributions and innovations proposed in this paper?” It extracted about 13,000 tokens. Let’s run this and observe the generation speed.
I encountered an error with AI Gemini 2.0 Flashlight, but after rerunning the same prompt, it took about 15 seconds, while AI Gemini 2.0 Flash took about 12 seconds. Now, let’s compare the results.
Both mention architecture, multi-head attention, Deep Seek with auxiliary loss, free load balancing, and multi-token predictions as main contributions.
Gemini 2.0 Flash highlighted auxiliary loss, free load balancing, and multi-token predictions, interestingly missing multi-head attention.
Both discuss training efficiency, mentioning FP8 mixed precision training. Gemini 2.0 Flash mentions the dual-pipe algorithm, efficient cross-node communication, and memory footprint optimization.
AI Gemini 2.0 Flashlight mentions FP8 code design of algorithms, frameworks, and hardware, and high-quality pre-training. It also discusses the dual-pipe algorithm and efficient cross-node communication.
Both retrieve similar information, and Gemini 2.0 Flash even recognized the MLA. Overall, the summaries are very similar, and either could be used.
Secret of Digital Word https://secretofdigital.world/