Unverified50% confidenceBenchmarkExact time
In Path A, an RTX 3060 runs the front layers and an AMD AI Max+ 395 runs the back layers, improving Prefill speed from 136.69 token/s to 319.1 token/s under IQ3 quantization with 64K long input.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/26/2026
First Seen
Valid until: 12/25/2026
Sources
Related Entities
Related Claims
UnverifiedIn terms of generation speed, Ollama's range is 339–448 tokens/s (peak 448), while OpenCode Go's range is 293–354 (peak 354).72% similarUnverifiedAt a 128K context, Q2_0 outputs 65.1 tok/s, IQ2_XS outputs 52.0 tok/s, and IQ3_XXS outputs 44.8 tok/s.67% similarUnverifiedQwen 3.8 has a base speed of 22 tokens/s on the M5 Max, which can be boosted to 34 or even 56 tokens/s with MTP enabled.67% similarUnverifiedTo complete the 250-case test, FP8 27B took 36 minutes and 27 seconds, while Flash-Next IQ4_XS took 2 hours, 5 minutes, and 47 seconds, a difference of approximately 3.5 times.61% similarUnverifiedIn token generation tests for the Gemma 4 12B model, Spark reached 26.4 tokens/s, Halo 24.6 tokens/s, and M4 Pro was the fastest at 33.8 tokens/s.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/952580API
curl https://kongchang.com/api/v1/knowledge/claims/952580MCP
get_claim(id=952580)