Eagle releases EAGLE-4: Qwen inference accelerated by 3.6 times, switches to a restrictive commercial license
BIBIBI
AT A GLANCE
Eagle releases EAGLE-4; specific tests in the report show improved generation speed and throughput, but its commercial license is more restrictive than those of the previous generations.
Article
EAGLE-4 Make Qwen generate faster by 3.6 times; smartphones CPU can also be accelerated by 3 times
Beating News:AI Eagle, a large-model inference acceleration company, has released EAGLE-4, upgrading its speculative decoding solution. The new version focuses on reducing the computational overhead of the draft model and can adjust the prediction scale based on hardware performance and concurrency levels, preventing additional prediction computation from instead slowing performance.
On 4 NVIDIA B300 GPU,Qwen3.8-27B's single-request generation speed increased from 240.5 Token per second to 863.8 Token, reaching 3.6 times that of standard inference. With 512 requests running simultaneously, total throughput increased from 21200 Token per second to 28300, an increase of 34%.
Mobile devices also saw significant gains. On the Motorola Razr 70 Ultra,Gemma-4-E2B-it achieved CPU times faster long-text generation using only 3.2, reducing the time for the same task from 43 seconds to 14 seconds and lowering energy consumption per Token by 53%。
EAGLE-4 has released its code publicly and supports integration with SGLang, vLLM, and experimental Token Speed. However, unlike the previous generations, which used the Apache 2.0 open-source license, the new version uses a restrictive commercial license. Enterprises with aggregate revenue exceeding 12 for 20000000 USD consecutive months, as well as companies providing competing inference services, require separate authorization for commercial use.
Original link https://m.theblockbeats.info/flash/372995
Key points
01
Eagle, a large-model inference acceleration company, has released EAGLE-4, upgrading its speculative decoding solution.
02
According to reports, the new version reduces the computational overhead of the draft model and can adjust the prediction scale based on hardware performance and concurrency levels.
03
On 4 NVIDIA B300 GPU,Qwen3.8-27B's single-request generation speed increased from 240.5 Token per second to 863.8 Token, 3.6 times that of standard inference.
04
With 512 requests running simultaneously, total throughput increased from 21200 Token per second to 28300 Token, an increase of 34%。
05
On the Motorola Razr 70 Ultra,Gemma-4-E2B-it achieved CPU using only 3.2, delivering 43 times faster long-text generation and reducing the time for the same task from 14 seconds to Token seconds, while lowering energy consumption per 53%。
06
EAGLE-4 has released its code publicly and supports integration with SGLang, vLLM, and experimental Token Speed。
07
Unlike the previous generations, which used the Apache 2.0 open-source license,EAGLE-4 uses a restrictive commercial license.
08
Enterprises with aggregate revenue exceeding 12 for 20000000 USD consecutive months, as well as companies providing competing inference services, require separate authorization for commercial use.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
EAGLE-4 is a speculative decoding solution: it uses a draft model to predict content in advance, which is then processed by a large model in an attempt to generate text faster. The reported GPU and mobile CPU tests both show speed improvements. The new version also adjusts the prediction scale according to hardware and concurrency conditions to reduce ineffective computation. However, publicly released code does not mean commercial use is unrestricted; some enterprises and inference service providers must obtain separate authorization.
Why it matters to readers
If the effects reported can be replicated in real-world deployments,EAGLE-4 it may increase single-request speed and concurrent throughput while reducing mobile generation time and energy consumption; however, the more restrictive commercial license will affect whether some enterprises can adopt it directly.
Beginners should distinguish key points: single-request speed indicates how fast one generation is; throughput indicates a system's ability to handle many requests simultaneously; and publicly available code does not necessarily mean that enterprises may use it commercially without restriction.
Risks and unknowns
The materials do not include an official Eagle announcement, a complete benchmarking methodology, or original test results.
The materials do not provide independent third-party replication, and it remains unclear whether the performance and energy-consumption data can be reproduced.
Related Developments
Loading event timeline…
The test results are based on specific models, hardware, and concurrency conditions and cannot be confirmed as generalizable to other deployment environments.
The materials only summarize the commercial licensing threshold and do not provide the complete license terms.
SGLang, vLLM, and experimental Token Speed's actual integration results are not explained in the materials.
Related concepts
Speculative decoding
This term is not in the glossary yet. Browse related concepts in the glossary.
The number of transactions a blockchain can process within a given period, an important indicator for assessing the capacity to support high-frequency AI activities.
