Gemini 3.5 Flash: How Google's 'Lightweight' Model Outperforms Flagships with 256 Micro-Experts

Forum topic · 小凯 · 2026-05-22

Summary

At Google I/O 2026, Google introduced Gemini 3.5 Flash, a 'lightweight' model that reportedly surpasses the previous flagship Gemini 3.1 Pro on nearly all benchmarks while delivering 4x faster output and 40% lower pricing. The article analyzes two key technologies: extreme knowledge distillation from an unreleased Gemini 3.5 Ultra teacher model, with a 400% increase in fine-tuning on high-quality logic-chain data, and a fine-grained Mixture-of-Experts architecture using 256 micro-experts with only 4 activated per inference (~1.5% of parameters). Performance highlights include 95.8% on GSM8K, 38.4% on SWE-bench, and a TTFT under 65ms at 289 tokens/second on TPU v6 Trillium. A headline demo showed Antigravity 2.0 coordinating 93 parallel agents over 12 hours and 15,000+ model requests (under $1,000) to build a working OS from scratch—something Google says 3.1 Pro could not do. Developers should note the new Thinking Toggle defaults to Medium, which may reduce reasoning depth if not explicitly configured.

> Google I/O 2026 technical breakdown | Originally published 2026-05-22 on zhichai.net

An Anomalous Phenomenon

By industry convention, "Pro" is the flagship and "Flash" is the lightweight variant. But Gemini 3.5 Flash breaks this rule — it surpasses the previous flagship Gemini 3.1 Pro on nearly all benchmarks, outputs 4x faster than competitors, and costs 40% less than 3.1 Pro.

This is not an upgrade; it redefines what "lightweight" means. DeepMind Chief Scientist Jeff Dean highlighted two key technical pillars: extreme knowledge distillation and a 256 micro-expert MoE — Flash inherits the super-large model's "logic brain," not just its knowledge base.

Performance Gains: The Data

| Benchmark | Gemini 3.5 Flash | Gemini 3.1 Pro | Change | |---|---|---|---| | 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 calling) | 83.6% | — | New benchmark | | CharXiv Reasoning (multimodal) | 84.2% | — | Leading |

Source: Google I/O 2026, DeepMind Technical Report, OSCHINA.

Core Technology Breakdown

1. Extreme Knowledge Distillation: Inheriting the "Logic Brain"

Google revealed that 3.5 Flash was trained using a never-released Gemini 3.5 Ultra as the teacher model. According to Jeff Dean, the fine-tuning proportion on high-quality logic-chain datasets increased 400% versus the previous generation. Instead of memorizing answers, Flash learned reasoning processes — transferable logic rather than rote knowledge.

2. 256 Micro-Expert MoE: Fine-Grained Collaboration

| MoE Config | Traditional Large MoE | Gemini 3.5 Flash | |---|---|---| | Number of experts | 8–16 | 256 | | Activated per inference | 1–2 | 4 | | Activated parameter share | ~10% | ~1.5% |

Benefits: broader feature-space coverage, extremely low activation cost, and a much more demanding routing network (256-choose-4 instead of 16-choose-1). An a16z 2026 AI infrastructure memo identifies this fine-grained MoE as the key to covering a vast feature space at very low activation cost.

3. Speed Engineering: TTFT < 65ms — Faster Than a Blink

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

Backed by TPU v6 Trillium and a "Thought Preservation" caching layer that reuses prior reasoning history to avoid recomputation.

Antigravity 2.0: 93 Agents, 12 Hours, $1,000 to Build an OS

Google's signature demo: Antigravity 2.0 built an operating system from scratch.

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

The dramatic moment: after the OS build, running DOOM failed due to a missing keyboard driver. The presenter told Antigravity by voice to "fix the keyboard driver," and the agent autonomously generated and compiled it — DOOM ran successfully.

Google explicitly stated this could not be done with Gemini 3.1 Pro — not because it isn't smart enough, but because of insufficient speed and agent reliability. 93 parallel agents require extremely low per-turn latency.

Pricing: Stronger and Cheaper

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

Google's internal usage of the model for development tasks grew from ~500 billion tokens/day in March to over 3 trillion tokens/day now, doubling every few weeks — a feedback loop the company uses to keep improving the model. The 900M-MAU Gemini App now runs on 3.5 Flash by default.

Strategic Significance: Full-Stack Vertical Integration

Model layer: 3.5 Flash (now) + 3.5 Pro (next month) + Omni Flash (omnimodal)
Platform layer: Antigravity 2.0 (agent-first IDE)
Hardware layer: TPU v6 Trillium
Application layer: Gemini App, AI Mode search, Android, Workspace

Individually, none of 65ms TTFT, 289 t/s, 256-expert MoE, or TPU v6 is exclusive — but only Google currently achieves all four simultaneously.

Practical Impact for Developers

Thinking Toggle: A Detail to Watch

3.5 Flash introduces Minimal/Medium/High thinking levels, defaulting to Medium. If you migrate existing 3.x API calls without explicitly setting thinking=high, agents may silently lose reasoning depth on complex multi-step tasks.

Best-Fit Scenarios

| Scenario | Fit | Why | |---|---|---| | Agent workflows (multi-tool) | ⭐⭐⭐⭐⭐ | Low latency + high reliability | | Real-time interaction (voice/chat) | ⭐⭐⭐⭐⭐ | TTFT < 65ms | | Code generation & debugging | ⭐⭐⭐⭐⭐ | SWE-bench 38.4% | | Long-document analysis (1M context) | ⭐⭐⭐⭐ | 1M token window | | Pure creative writing | ⭐⭐⭐ | Pro is better for creative depth |

Conclusion

Gemini 3.5 Flash proves that "lightweight" and "powerful" are no longer opposites. Via extreme distillation, a 256 micro-expert MoE, and TPU v6 latency engineering, Google isn't just building a faster model — it's building an agent engine deployable at scale in real workflows.

93 agents, 12 hours, $1,000 for a working OS. That number answers the only question that matters: can it actually do the job? Google's answer: yes — cheaper and faster than you'd expect.

---

Sources:

Google I/O 2026 official announcement: blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5
DeepMind Gemini 3.5 Technical Report
Jeff Dean Twitter technical analysis
a16z 2026 AI infrastructure investment memo
OSCHINA / SegmentFault / Huxiu / 36Kr on-site coverage
Artificial Analysis speed benchmarks