Gemini 3.5 Flash: Redefining Speed and Efficiency with 256 Micro-Experts

Forum topic · 小凯 · 2026-05-22

Summary

At Google I/O 2026, Google unveiled Gemini 3.5 Flash, a lightweight model that surpasses its predecessor Gemini 3.1 Pro on most benchmarks while running 4x faster and costing 40% less. The model combines extreme knowledge distillation from an unreleased Gemini 3.5 Ultra teacher with a fine-grained 256-expert Mixture-of-Experts architecture that activates only 4 experts per inference (~1.5% of total parameters). Key results include 95.8% on GSM8K, 38.4% on SWE-bench, 289 tokens/sec output, and a time-to-first-token under 65ms, faster than a human blink. Powered by TPU v6 Trillium, Flash enabled the Antigravity 2.0 demo where 93 parallel agents built an OS in 12 hours for under $1,000. Google now serves 9 billion daily tokens through Gemini products, with 900 million MAUs on the Gemini App running on Flash by default.

Key points

Performance breakthrough

Gemini 3.5 Flash outperforms the previous flagship 3.1 Pro across nearly every benchmark while being 4x faster and 40% cheaper. Notable results:

| Benchmark | 3.5 Flash | 3.1 Pro | Delta | |-----------|-----------|---------|-------| | GSM8K (math reasoning) | 95.8% | 93.2% | +2.6% | | SWE-bench (code generation) | 38.4% | 32.1% | +6.3% | | Terminal-Bench 2.1 | 76.2% | 70.3% | +5.9% | | GDPval-AA Elo | 1656 | 1314 | +26% | | MCP Atlas (agent tool use) | 83.6% | — | new high | | CharXiv Reasoning (multimodal) | 84.2% | — | leading |

Core technical innovations

1. Extreme knowledge distillation. The student model is distilled from an unreleased Gemini 3.5 Ultra teacher. Per Jeff Dean, fine-tuning on high-quality chain-of-thought data rose 400% versus the prior generation. The student learns *reasoning processes*, not just answers, yielding transferable logic instead of memorized knowledge.

2. 256 micro-expert MoE. The model uses 256 experts and activates only 4 per inference:

| MoE config | Traditional large MoE | 3.5 Flash | |------------|----------------------|-----------| | Number of experts | 8–16 | 256 | | Activated per step | 1–2 | 4 | | Active parameter share | ~10% | ~1.5% |

This fine-grained routing covers a broader feature space (e.g., one expert for recursive algorithms, another for geometric reasoning) while keeping compute costs minimal. Routing becomes a precision scheduler.

3. Speed engineering.

| Metric | 3.5 Flash | Comparable competitor | |--------|-----------|----------------------| | Output speed | 289 tokens/sec | Claude Opus 4.7: 67 t/s | | TTFT (first-token latency) | <65 ms | Human blink: 100–150 ms | | After Antigravity optimization | 12x competitor speed | — |

TTFT under 65 ms means the model starts producing output before a user finishes blinking. A "Thought Preservation" caching layer reuses prior inference history. Hardware: TPU v6 Trillium.

Antigravity 2.0 demo

The flagship demonstration had Antigravity 2.0 build an operating system from scratch:

| Metric | Value | |--------|-------| | Parallel sub-agents | 93 | | Duration | 12 hours | | Model requests | 15,000+ | | Tokens processed | 2.6 billion | | Total API cost | <$1,000 |

The OS successfully ran DOOM after agents generated and compiled a missing keyboard driver via voice prompt. Google confirmed this was impossible on 3.1 Pro due to latency and agent reliability limits.

Pricing and scale

| Model | Input ( $$/M tokens) | Output ($$ /M tokens) | |-------|--------------------|---------------------| | Gemini 3.5 Flash | $1.50 | $9.00 | | Gemini 3.1 Pro (approx.) | $2.50 | $15.00 |

Google now processes over 3 trillion tokens per day for internal dev tasks (up from 500 billion in March), with the Gemini App's 900 million MAUs running on Flash by default.

Strategic stack

Google's vertical integration spans model (Flash today, Pro next month, Omni Flash multimodal), platform (Antigravity 2.0), hardware (TPU v6 Trillium), and applications (Gemini App, AI Mode search, Android, Workspace). Combining 65 ms TTFT, 289 t/s, 256-expert MoE, and TPU v6 simultaneously is unique to Google today.

Developer notes

Flash introduces Minimal/Medium/High Thinking controls, default Medium. Migrating older API calls without explicitly setting thinking=high may reduce reasoning depth for complex multi-step tasks.
Best fit: agentic workflows with frequent tool calls, real-time voice/chat, code generation, long-document analysis (1M context). Pure creative writing may still favor Pro.

Conclusion

Gemini 3.5 Flash demonstrates that "lightweight" and "powerful" are no longer opposites. Through distillation of reasoning (not just answers), extreme MoE sparsity, and sub-blink latency, Google has created a deployable agent intelligence engine at scale.

Sources: Google I/O 2026 official blog (blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5), DeepMind Gemini 3.5 Technical Report, Jeff Dean Twitter analysis, a16z 2026 AI Infrastructure Memo, OSCHINA / SegmentFault / Huxiu / 36Kr live reports, Artificial Analysis benchmarks.