Unverified50% confidenceEventExact time
A developer running the Qwen3 27B model on an M5 Pro/64GB saw a first token of only 6.1 seconds on a cache hit within the same session, but after the cache became invalid, re-reading about 41,000 tokens caused a first-token latency of 268 seconds, a 44-fold difference in speed.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/1/2026
First Seen
Valid until: 1/30/2027
Sources
Related Entities
Related Claims
UnverifiedRunning via llama.cpp on a MacBook Pro with an M4 Pro chip and 24GB unified memory, the text decoding speed is 10.13 tokens/second, with each complete task taking 4 to 7 minutes.79% similarUnverifieddecode阶段的生成速度近似等于内存带宽÷模型大小，例如在内存带宽273GB/s的M4 Pro上运行4GB的Q4量化7B模型理论最大生成速度约68 tokens/s79% similarUnverifiedOn the same machine and model, the decoding speed difference between low power mode and high power mode is 1.4 to 3.8 times; for example, running Qwen 3.5 35B on Ollama gives 51.9 tokens/s in low power and 107.1 tokens/s in high power.74% similarUnverifiedWhen running Qwen3-Next-Flash (Autoround W4A16 quantization), the prefill speed was approximately 1.3k tokens/s, and the generation speed reached 70 tokens/s (code) / 60 tokens/s (prose), with support for 128k+ long context.74% similarUnverifiedSwitching the model to LFM2.5-2.6B (run via the Mirai AI server), on the same OpenClaw task, reduced memory usage to 17GB with a generation speed of about 54.2 tokens/second, and the task was not completed in time.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1122340API
curl https://kongchang.com/api/v1/knowledge/claims/1122340MCP
get_claim(id=1122340)