|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Llada的研究人員報告了他們的80億參數模型的性能與各種基準的Llama3 8b相似

Large language models (LLMs) are typically bottlenecked by speed, especially when compared to smaller models. However, new diffusion-based models have reportedly achieved performance faster than or comparable to similarly sized conventional models while also being significantly faster.
大型語言模型(LLM)通常是通過速度瓶頸的,尤其是與較小的型號相比。但是,據報導,基於擴散的新模型的性能比相似大小的常規模型更快或相當,同時也要快得多。
LLaDA, a 7-billion parameter language model from researchers at the University of Science and Technology of China, performs similarly to LLaMA3 8B across various benchmarks.
蘭達(Llada)是中國科學技術研究人員的70億個參數語言模型,在各種基準測試中與Llama3 8B的表現相似。
According to Inception, an AI startup focused on efficient deep learning, LLaDA achieves competitive results on tasks like MMLU, ARC, and GSM8K.
根據Inception,一家專注於有效深度學習的AI初創公司在MMLU,ARC和GSM8K等任務上取得了競爭成果。
But when comparing the speed of an 8-billion parameter LLaDA model to an 8-billion parameter LLaMA3 model, LLaDA generates 1,024 tokens per second on an Nvidia H100, whereas LLaMA3 generates about 60 tokens per second. This difference in throughput is roughly 17x.
但是,將80億個參數LLADA模型與80億參數Llama3模型進行比較時,LLADA在NVIDIA H100上每秒產生1,024個令牌,而Llama3每秒產生約60個令牌。吞吐量的這種差異約為17倍。
LLaDA's researchers claim this level of performance is possible with models as small as 3 billion parameters.
LLADA的研究人員聲稱,這種績效水平是可能的,模型小於30億個參數。
Compared to GPT-4o Mini, which operates at 59 tokens per second, Mercury Coder Mini scores 88.0 percent on HumanEval and 77.1 percent on MBPP—comparable to GPT-4o Mini—while reportedly operating at 1,109 tokens per second. This represents an 18x speed advantage over GPT-4o Mini.
與GPT-4O MINI相比,該Mini的運行量為每秒59個令牌,汞編碼器Mini在HOMANEVAL上得分為88.0%,MBPP的分數為77.1%,與GPT-4O Mini相吻合,據報導以每秒1,109個代幣運行。這代表了比GPT-4O Mini具有18倍的速度優勢。
Mercury's documentation states its models run "at over 1,000 tokens/sec on Nvidia H100s, a speed previously possible only using custom chips" from specialized hardware providers like Groq, Cerebras, and SambaNova. When compared to other speed-optimized models, the claimed advantage remains significant—Mercury Coder Mini is reportedly about 5.5x faster than Gemini 2.0 Flash-Lite (201 tokens/second) and 18x faster than Claude 3.5 Haiku (61 tokens/second).
水星的文檔指出,其模型“在NVIDIA H100上以1000多個令牌/秒為單位,以前只能使用定制芯片從Groq,Cerebras和Sambanova等專業硬件提供商使用自定義芯片”。與其他速度優化模型相比,聲稱的優勢仍然顯著 - 據報導,MINI的MINI比Gemini 2.0 Flash-Lite(201/秒/秒)快約5.5倍,比Claude 3.5 Haiku(61個令牌/秒)快18倍。
According to Inception, Mercury Coder Mini also outperforms smaller models from GPT-3.5 series and PaLM 2 series at comparable speed.
根據Inception,水星編碼器Mini還以可比的速度優於GPT-3.5系列和Palm 2系列的較小型號。
At 1,024 tokens per second, LLaDA is faster than smaller models from GPT-3.5 series and PaLM 2 series at comparable speed. At 201 tokens per second, Gemini 2.0 Flash-Lite has better performance at HumanEval and worse performance at MBPP than Mercury Coder Mini. At 61 tokens per second, Claude 3.5 Haiku has better performance at both HumanEval and MBPP than Mercury Coder Mini.
LLADA在每秒1,024個令牌時的速度比GPT-3.5系列和Palm 2系列的較小型號以可比的速度快。 Gemini 2.0 Flash-lite在每秒的201個令牌上,在HOMANEVAL上的性能要好,而MBPP的性能比Mercury Coder Mini更差。 Claude 3.5 Haiku在每秒61個令牌時,在HumaneVal和MBPP的性能都比Mercury Coder Mini更好。
Opening a potential new frontier in LLMs
在LLMS開設潛在的新邊界
Diffusion models do involve some trade-offs. They typically need multiple forward passes through the network to generate a complete response, unlike traditional models that need just one pass per token. However, because diffusion models process all tokens in parallel, they achieve higher throughput despite this overhead.
擴散模型確實涉及一些權衡。他們通常需要多個向前通過網絡來產生完整的響應,這與傳統模型只需要一個令牌,而傳統模型也只需一個代幣。但是,由於擴散模型並行處理所有令牌,但儘管開銷,它們達到了更高的吞吐量。
Inception thinks the speed advantages could impact code completion tools where instant response may affect developer productivity, conversational AI applications, resource-limited environments like mobile applications, and AI agents that need to respond quickly.
Inception認為,速度優勢可能會影響代碼完成工具,即即時響應可能會影響開發人員生產力,對話AI應用程序,資源有限的環境(例如移動應用程序)以及需要快速響應的AI代理。
If diffusion-based language models maintain quality while improving speed, they might change how AI text generation develops. So far, AI researchers have been open to new approaches.
如果基於擴散的語言模型在提高速度的同時保持質量,則可能會改變AI文本生成的發展方式。到目前為止,AI研究人員已經接受了新方法。
Independent AI researcher Simon Willison told Ars Technica, "I love that people are experimenting with alternative architectures to transformers, it's yet another illustration of how much of the space of LLMs we haven't even started to explore yet."
獨立的AI研究員西蒙·威利森(Simon Willison)告訴ARS Technica:“我喜歡人們正在嘗試使用替代架構到變形金剛,這是我們甚至還沒有開始探索的LLMS空間的另一個說明。”
On X, former OpenAI researcher Andrej Karpathy wrote about Inception, "This model has the potential to be different, and possibly showcase new, unique psychology, or new strengths and weaknesses. I encourage people to try it out!"
在X上,前Openai研究員Andrej Karpathy寫道:“該模型有可能與眾不同,可能展示新的,獨特的心理學或新的優勢和劣勢。我鼓勵人們嘗試一下!”
Questions remain about whether larger diffusion models can match the performance of models like GPT-4o and Claude 3.7 Sonnet, and if the approach can handle increasingly complex simulated reasoning tasks. For now, these models offer an alternative for smaller AI language models that doesn't seem to sacrifice capability for speed.
關於較大的擴散模型是否可以與GPT-4O和Claude 3.7十四行詩等模型的性能相匹配的問題仍然存在問題,並且該方法是否可以處理日益複雜的模擬推理任務。目前,這些模型為較小的AI語言模型提供了一種替代方案,似乎並沒有為速度犧牲能力。
You can try Mercury Coder yourself on Inception's demo site, and you can download code for LLaDA or try a demo on Hugging Face.
您可以在Inception的演示網站上嘗試Mercury編碼器,並且可以下載llada的代碼,也可以在擁抱面孔時嘗試演示。
免責聲明:info@kdj.com
所提供的資訊並非交易建議。 kDJ.com對任何基於本文提供的資訊進行的投資不承擔任何責任。加密貨幣波動性較大,建議您充分研究後謹慎投資!
如果您認為本網站使用的內容侵犯了您的版權,請立即聯絡我們(info@kdj.com),我們將及時刪除。
-
-
-
-
-
-
- 比特幣零售需求因價格下跌而暴跌,技術指標確認看跌趨勢
- 2026-11-14 15:45:01
- Bitcoin零售需求銳減近10%至3月以來最低水平,伴隨價格跌破7萬美元,技術指標指向持續看跌。
-
-
-
- XRP 價格預測:1000 美元會讓大家變得富有嗎?技術面建議值得關注的關鍵水平
- 2026-11-14 04:05:01
- 分析 XRP 以 1,000 美元創造百萬富翁的潛力以及價格預測的關鍵技術水平。

































