Gemini 3.5 Flash on Mac: What the cloud model does — and what runs locally

Update October 11, 2026: Google marks gemini-3.5-flash as deprecated and automatically routes requests to gemini-3.6-flash, according to the model card updated October 9. For new Mac projects, use gemini-3.6-flash or gemini-3.8-flash directly. Both run in the cloud, unlike downloaded Gemma 4. Google status. See the separate API migration guide.

Model status: Gemini 3.5 Flash is deprecated and automatically routed to 3.6 Flash. For a new API integration, compare Gemini 3.8 Flash pricing before selecting the older model.¹

Gemini 3.5 Flash is a cloud API model with a documented 1,048,576-token input context and multimodal input; it is not an Ollama model you can download. This guide separates its supported API features, current costs and data retention from genuinely local Mac alternatives.²¹

Local execution: no. Google’s public documentation does not provide downloadable Gemini 3.5 Flash weights or an official GGUF/MLX build. A Mac app can call the Gemini API, but the inference takes place on remote infrastructure.²

Related guides: Gemini 3.5 Flash API deep dive, Ollama on Mac mini M4, LM Studio vs Ollama, Unified Memory on Mac, Gemma 4 on Mac and the open-weight shortlist.

What Gemini 3.5 Flash actually is

Gemini 3.5 Flash is deprecated; its legacy model ID routes to 3.6 Flash. The former model ID gemini-3.5-flash now routes to gemini-3.6-flash. Provider prefixes are routing metadata, not locally downloadable model weights.²

According to Google’s model page, it accepts text, image, video, audio and PDF input and returns text output. The input context window is 1,048,576 tokens with an output limit of 65,536 tokens.²

Feature-wise it supports thinking, function calling, code execution, Computer Use (Preview), File Search, URL context, Search Grounding, Maps Grounding, the Batch API, caching, Flex inference and Priority inference. Audio generation, image generation and the Live API are not supported.²

Google documents the Interactions API for new integrations, while generateContent remains available. The APIs are not feature-identical: Google’s current overview lists Batch requests and explicit caching under generateContent, not the Interactions API. Interactions are stored by default unless you set store=false.³⁴

How to use it from a Mac

There is nothing Mac-specific to install. You create an API key in Google AI Studio and call the model from any HTTP client, SDK or app that speaks the Gemini API:

Current model ID: gemini-3.6-flash

Example: Make an actual API request from your Mac

Google documents the Interactions API under /v1beta/interactions. Create your API key in Google AI Studio and make it available as the GEMINI_API_KEY environment variable. Do not commit the key to a repository or embed it in browser JavaScript. This call is remote and can incur charges in a paid project:

# Prerequisite: GEMINI_API_KEY is set in your current shell.
curl --fail-with-body -sS -X POST \
  'https://generativelanguage.googleapis.com/v1beta/interactions' \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gemini-3.6-flash",
    "store": false,
    "input": "Answer in one sentence: What is unified memory?"
  }'

According to Google, store: false prevents Interaction-object storage; it does not disable all abuse-monitoring logs or other processing. It also prevents server-side chaining and background execution. Your Mac sends the prompt to Google’s servers: this is not offline inference.⁴⁵⁶

Actually local: Gemma 4 instead of Gemini

Google offers Gemma 4, a newer open-weight generation than Gemma 3. Ollama lists downloadable gemma4:e2b, gemma4:e4b, gemma4:12b, gemma4:26b and gemma4:31b tags. For a small initial Mac test:

ollama pull gemma4:e2b
ollama show gemma4:e2b
ollama run gemma4:e2b

In contrast, gemma4:cloud runs remotely. Check the exact model tag and ollama ps, and disable unrelated browser, web-search and MCP cloud tools if your full workflow must stay local. Gemma 4 is not the same model as Gemini 3.5 Flash.⁷⁸⁹¹⁰

In practice, three access patterns matter for Mac users:

• Scripts and apps: call the REST endpoint or an official SDK with your key. Choose Interactions for its supported stateful or stateless flows. Use generateContent where the needed API capability—such as Batch or explicit caching—is not yet supported in Interactions.³⁴
• Agent tools: coding agents and multi-model routers commonly expose the model under provider naming such as google/gemini-3.5-flash. That prefix is routing metadata, not a different model.¹¹
• Google AI Studio: AI Studio lets you test API models available to your project. Model selection and billing depend on the account and project.²

If you want the full request/response mechanics, token accounting and code samples, see the dedicated Gemini 3.5 Flash API article.

What it costs

The legacy 3.5 model ID automatically routes to Gemini 3.6 Flash. Consult current 3.6 Flash rates, not the original 3.5 price sheet. According to Google, paid-tier USD per million tokens through December 31, 2026:

Mode	Input	Output incl. thinking
Standard	$0.75	$3.75
Batch / Flex	$0.375	$1.875
Priority	$1.35	$6.75

At 1M input plus 200,000 output tokens, 3.6 Flash and 3.8 Flash both cost $1.50 at current Standard list rates, excluding extras. Historical 3.5 prices of $1.50 input / $9.00 output, which imply $3.30 for that example, are not current billing rates for the deprecated alias. Google lists 3.6/3.8 Standard at $1.50/$7.50 from January 1, 2027. Grounding, cache storage and tools can incur extra charges. Matching rates do not imply equal quality.

Privacy: what happens to your data

There are three different questions: Is the inference executed remotely? May a provider use submitted content to improve its models? How long are server-side interactions or logs retained? Answer them separately.¹⁴⁵⁶

• Inference location: Gemini 3.5 Flash runs through Google’s API, not on your Mac. Ollama Cloud likewise processes its cloud-model requests remotely. Downloaded Ollama models can run locally if your client and tools do not forward data elsewhere.²⁷⁸
• Product improvement: Google’s Gemini API pricing table marks Free-tier content as eligible for product improvement and paid-tier content as not used for that purpose. This statement is not a promise of zero logging or zero retention.¹
• Stored Interactions: The Interactions API defaults to store=true. Google documents 55-day default retention on paid tier and one day on free tier. Paid-tier projects can select 7, 14, 28 or 55 days in AI Studio. To avoid storing an Interaction object, send store=false; that disables server-side conversation chaining with previous_interaction_id and cannot be combined with background execution.⁴
• Separate policy controls: Logging and abuse-monitoring policies may apply independently of Interaction storage. Check the current Google documentation, your project settings and any contractual obligations rather than treating store=false as a universal no-retention switch.⁵⁶
• Ollama Cloud: Ollama documents that it processes cloud prompts and responses to deliver the service and does not use them for model training. Its local/cloud distinction is separate from Google’s Gemini API.⁷

For sensitive or regulated information, verify all data paths, client telemetry, API storage settings and applicable provider terms before uploading anything.

Does Gemini 3.5 Flash run locally in Ollama, LM Studio or MLX?

No. The official sources reviewed here contain no Gemini 3.5 Flash weights, GGUF or MLX path, and no Apple-Silicon runtime. A command like ollama run gemini-3.5-flash does not start this model on your Mac.²¹¹¹²

Two related traps are worth knowing:

• Provider naming is not a download. google/gemini-3.5-flash in routers or benchmark tools means provider Google, model Gemini 3.5 Flash — it is not an Ollama or local weight path.¹¹
• Ollama Cloud is not local inference. Ollama documents Cloud Models that run in Ollama’s infrastructure and require an account. Useful when your Mac lacks memory, but your data leaves the machine.⁷

Rule of thumb: an Ollama command does not automatically mean local inference. With cloud models, Ollama acts as a local interface while inference runs elsewhere.

This is a time-bound documentation statement, not proof about every possible community project. If you see a Gemini-style entry in a local tool, check whether it is a different model, a preview/provider/cloud entry, or something that sends your prompt to a remote service.²¹¹

Local alternatives when you need offline AI

If what you want is “this class of capability, but running on my Mac”, there is no 1:1 replacement — you pick an open-weight model that fits your hardware:¹³¹⁴¹²

Goal	Local direction	Note
Google open-weight models	Gemma 4 (e.g. E2B, E4B, 12B)	newer local generation than Gemma 3; not Gemini; confirm tag and RAM
General chat	Qwen / Llama / Mistral class	Depends on size and quantization
Coding	Qwen / DeepSeek / code-focused models	Quality varies strongly with size
Vision	Gemma 4 or another vision-capable model	Check modality on the exact Ollama tag
Transcription	Whisper	Different family, very practical locally
Private documents	Local RAG + Ollama/LM Studio	Privacy benefit only in a local-only setup

Rough Mac RAM orientation — heuristics, not benchmarks:¹⁵

• 8 GB: small 1B–4B models, sometimes heavily quantized 7B.
• 16 GB: 7B/8B comfortably, some 12B workflows.
• 24 GB: 12B/14B realistic, larger models carefully.
• 32 GB: 27B-class becomes realistic with limited context.
• 48 GB+: larger models and vision workflows get comfortable.

Runtime, quantization, context length, KV cache, swap and open apps change real memory use. For the details, see how much RAM local LLMs need.

A three-step Mac setup check

1. Choose the intended execution location. For Gemini, send the official model ID gemini-3.5-flash to Google. For offline work, download an open-weight model and confirm that the inference engine is running locally.
2. Check actual cost and API behavior. Compare Google Standard prices against 3.8 Flash’s temporary rates; pick Interactions or generateContent based on which capabilities the application needs. Count thinking output and tool charges separately.¹⁴
3. Verify the data path. For hosted requests, inspect storage and provider settings. For local requests, verify the model is loaded locally and that no browser plugin, external tool or cloud relay silently forwards the content.⁷⁸

A Google Gemini model ID is not a download tag. For truly offline inference, start with a compatible Gemma, Qwen or other open-weight package instead of looking for a fictitious Gemini 3.5 GGUF file.

Sources

Footnotes

1. https://ai.google.dev/gemini-api/docs/pricing ↩ ↩² ↩³ ↩⁴ ↩⁵

2. https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

3. https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.5 ↩ ↩²

4. https://ai.google.dev/gemini-api/docs/interactions ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

5. https://ai.google.dev/gemini-api/docs/logs-policy ↩ ↩² ↩³

6. https://ai.google.dev/gemini-api/docs/usage-policies ↩ ↩² ↩³

7. https://docs.ollama.com/cloud ↩ ↩² ↩³ ↩⁴ ↩⁵

8. https://docs.ollama.com/faq ↩ ↩² ↩³

9. https://ollama.com/library/gemma4 ↩

10. https://ai.google.dev/gemma/docs/core/model_card_4 ↩

11. https://ollama.com/search?q=gemini ↩ ↩² ↩³ ↩⁴

12. https://github.com/ml-explore/mlx ↩ ↩²

13. https://ollama.com/library/gemma3 ↩

14. https://www.lmstudio.ai/docs/app/offline ↩

15. Derived from public model card sizes and community reports; ai-on-mac ran no own benchmark here. ↩