⚛️量子位•較早收集於 27m
擴散模型達每秒1009 tokens最快深度思考
#diffusion-models#inference-speed#llm-generationdiffusion-modelsnvidiamicrosoft
💡Diffusion at 1009 tps beats autoregressive – Nvidia/MS invest, paradigm shift incoming!
⚡ 30-Second TL;DR
有什麼變化
擴散模型生成每秒1009個tokens
為什麼重要
此突破可能終結對緩慢自迴歸生成的依賴，讓LLM實現即時複雜推理。投資顯示產業轉向擴散模型以擴展推理。
下一步行動
Implement diffusion-based generation from arXiv papers like Diffusion-LM to benchmark against autoregressive baselines.
誰應關注：Researchers & Academics
關鍵要點
- •擴散模型生成每秒1009個tokens
- •速度超越自迴歸模型
- •英偉達與微軟投資
- •AI應如編輯而非打字機
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 8 個來源。
🔑 增強重點摘要
- •Inception Labs, founded by Stanford professor Stefano Ermon who contributed to Stable Diffusion and DALL-E, released Mercury 2 on February 24, 2026, backed by a $50M seed round from Menlo Ventures.[1][5]
- •Mercury 2 features a 128,000 token context window, native tool use, tunable inference, and schema-compliant JSON output, targeting latency-sensitive applications like voice AI, code editors, and agentic loops.[3][4][5]
- •The model achieves end-to-end latency of 1.7 seconds, undercutting competitors on benchmarks such as AIME 2025 (91.1), GPQA (73.6), and LiveCodeBench (67.3).[6][7]
📊 競品分析▸ Show
| Model | Speed (tokens/sec) | Input Price ($/M) | Output Price ($/M) | Latency (sec) | Context Length |
|---|---|---|---|---|---|
| Mercury 2 | 1009 | 0.25 | 0.75 | 1.7 | 128K |
| Claude Haiku 4.5 | ~89 | 1.00 | 5.00 | 23.4 | N/A |
| GPT-5 Mini | ~71 | N/A | N/A | N/A | N/A |
| Gemini 3 Flash | N/A | 0.50 | 3.00 | 14.4 | N/A |
🛠️ 技術深入
- •Replaces autoregressive sequential decoding with diffusion-based parallel refinement: starts with noisy sketch of response and denoises over few steps to coherent text.[1][5][6]
- •Processes entire sequences simultaneously using bidirectional attention, enabling multiple tokens at once but optimized to reduce recomputation costs per step.[1]
- •Runs on NVIDIA Blackwell GPUs; supports 128K context, native tools, tunable reasoning depth, and schema-aligned JSON mode.[3][4][5]
🔮 前景展望AI analysis grounded in cited sources
Diffusion LLMs will capture >30% of latency-critical inference market by 2027
NVIDIA Blackwell adoption surges 50% for text diffusion models in 2026
⏳ 時間線
2024-12
Inception Labs founded by Stanford's Stefano Ermon, leveraging diffusion expertise from Stable Diffusion/DALL-E work.
2025-01
Inception secures $50M seed funding from Menlo Ventures to commercialize diffusion for language models.
2026-02
Mercury 2 released on February 24 as first production diffusion reasoning LLM, hitting 1009 tokens/sec on Blackwell GPUs.
📎 來源 (8)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- implicator.ai — Inception Ships Mercury 2 a Diffusion LLM That Hits 1 009 Tokens Per Second
- techmeme.com — P22
- gigazine.net — 20260225 Inception Mercury 2
- deeplearning.ai — Anthropic U S Square Off Over AI Safeguards
- inceptionlabs.ai — Introducing Mercury 2
- the-decoder.com — Inception Launches Mercury 2 the First Diffusion Based Language Reasoning Model
- businesswire.com — Inception Launches Mercury 2 the Fastest Reasoning LLM 5x Faster Than Leading Speed Optimized Llms with Dramatically Lower Inference Cost
- perplexity.ai — Inception Launches Mercury 2 a Xwostpaurasjoadyw65o7g
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： 量子位 ↗
每週 AI 簡報
每週一封，可隨時退訂。