Gemini 3.7 Flash Model Card

TL;DR

Google DeepMind’s official model card for Gemini 3.7 Flash. Positioned as the next iteration in the Gemini 3 family with “algorithmic improvements to its core reasoning foundation” over 3.6 Flash — same architecture, training dataset, hardware, and software as 3.6 Flash (deferred to that model card), same 1M-token input context and 64K-token output, same input modalities (text, image, audio, video). Ships in Gemini App Spark, Gemini Enterprise, AI Studio, Gemini API, and Antigravity. Introductory price expires 2026-12-31 (then $1.50/1M input,$ 7.50/1M output). Knowledge cutoff March 2026 for some domains, January 2025 for others. Frontier Safety assessment reports it did not reach any tracked or critical capability levels; separate FSF report deferred.

Key claims

Gemini 3.7 Flash is built on Gemini 3.6 Flash — architecture, training dataset, training-data processing, hardware, and software are all delegated to the 3.6 Flash model card [§Model Information, §Model Data, §Implementation and Sustainability].
The advertised improvement over 3.6 Flash is framed as “algorithmic improvements to its core reasoning foundation” plus customizable thinking configurations for the quality/cost/latency mix [§Description].
Input modalities are text, images, audio, and video, with a 1M-token context window; output is text with a 64K-token cap [§Inputs, §Outputs].
Distribution channels are Gemini App Spark, Gemini Enterprise App, Gemini Enterprise Agent Platform, Google AI Studio, Gemini API, and Google Antigravity [§Distribution].
Introductory price expires 2026-12-31, after which $1.50/1M input tokens and$ 7.50/1M output tokens apply [§Evaluation Results footnote].
Knowledge cutoff is March 2026 for some domains; users may see the model’s knowledge limited to January 2025 in others, in line with the Gemini 3 Model Family [§Known Limitations].
Safety performance is reported as “similar to Gemini 3.6 Flash across both safety and tone, with low unjustified refusals,” measured with an updated internal evaluation set that is not directly comparable to numbers in earlier Gemini model cards [§Training and Development Evaluation Results].
Under Google’s Frontier Safety Framework (April 2026 revision), Gemini 3.7 Flash did not reach any tracked or critical capability levels; the full 3.7 FSF report is promised as a separate publication [§Frontier Safety Assessment].
The release ships with updated safeguards for CBRN and cyber-offense misuse [§Frontier Safety Assessment].
Numeric benchmark results are referenced as “listed below” but are delivered as images / an external methodology page (deepmind.com/models/evals-methodology/gemini-3-7-flash), not as inline numbers in the card text [§Evaluation Results].

Method

The card itself does not describe method. Every technical section (architecture, training data, training data processing, hardware, software, ethics evaluation approach, safety policies, acceptable usage) is deferred to the Gemini 3.6 Flash model card. The only 3.7-specific claims in the card body are (a) the words “algorithmic improvements to its core reasoning foundation,” (b) support for customizable thinking configurations, (c) the March 2026 knowledge-cutoff bump, and (d) an updated CBRN/cyber safeguard set. No training-compute number, no parameter count, no architecture change, no training-data delta, no distillation-source disclosure.

Results

Benchmark evaluation approach lists reasoning, coding, agentic tool use, multimodal capabilities, multi-lingual performance, and long-context, with methodology linked to deepmind.com/models/evals-methodology/gemini-3-7-flash [§Approach]. The results themselves are delivered visually (image-only in the fetched card) and are not machine-readable from the model-card body. The one concrete numeric commitment is pricing: introductory expires 2026-12-31, then $1.50/1M input tokens and$ 7.50/1M output tokens.

Safety-eval delta vs 3.6 Flash is reported as “similar across both safety and tone, with low unjustified refusals,” with a positive-percentage-change color-coded table (not machine-readable here) and a note that the evaluation methodology was refined to reduce false positives/negatives — meaning results are not comparable to earlier Gemini model-card numbers [§Training and Development Evaluation Results].

Human red-teaming reports: passing child-safety launch thresholds, “similar or improved” content-safety performance vs 3.6 Flash, and no egregious concerns relative to Gemini 3.1 Pro [§Human Red Teaming Results].

Why it’s interesting

Sits inside the tight Gemini 3 Flash release cadence — Build with Gemini 3 Flash, frontier intelligence that scales with you launched the base 3 Flash in December 2025 at roughly 1/4 the cost of 3 Pro, and Google has since shipped 3.5 → 3.6 → 3.7 at roughly three-week intervals, with 3.7 explicitly framed as an algorithmic improvement over 3.6 rather than a scale-up. Two implications worth flagging. First, the model card as a research artifact has been almost fully evacuated: architecture, training data, hardware, software, and evaluation methodology are all pointers rather than content, and the benchmark results are image-only. This is much thinner than the paired tech reports open frontier labs ship (e.g. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities Gemini 2.5, GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models GLM-4.5, Advancing Open-source World Models (LingBot-World) LingBot-World) and contrasts sharply with the open-release packaging pattern tracked under Open foundation-model releases. Second, the “algorithmic improvements to the reasoning foundation over 3.6” language matches the framing used by Introducing GPT-5.2 and GPT-5.3 Instant: Smoother, more useful everyday conversations — closed frontier labs increasingly ship incremental post-training / distillation refreshes to API tiers between architecture-changing releases, and Gemini 3.7 Flash is the sharpest example on the wiki so far (three weeks between 3.6 and 3.7). The companion tweet is filed as Gemini 3.7 Flash — 50% cheaper than 3.6, ~3 weeks between releases; this page is the model-card artifact behind that announcement, with the delta being (a) the CBRN/cyber safeguard update, (b) the customizable-thinking-configurations detail, and (c) the pricing schedule.