Uploaded September 2026 | Updated November 2026, 20 hours ago
MTP made Qwen 3.8 more than 2 times faster. But on essay writing, it made Ornith slower.
MTP means multi-token prediction. A small draft head guesses the next few tokens, and the full model checks those guesses in one go.
▸ Guess right → you get extra tokens
▸ Guess wrong → that compute is wasted
▸ On JSON: Qwen 3.8 got 2.3× the speed with MTP, Ornith 35B got only 1.1×
▸ On essays: Ornith 35B dropped to 0.7×, so it became slower
▸ Why? On essays, the guesses are right only around 30% of the time
So test MTP on your own content first.
Will you test MTP? Tell me in the comments 👇
Follow for more AI benchmarks.
#MTP #LLM #AI #Shorts #Qwen #LocalLLM #AIBenchmark
Full benchmark: kgptalkie.com/tutorials/llm-benchmarking/ornith-1-5-9b-vs-35b-a3b-benchmark
MTP made Qwen 3.8 more than 2 times faster. But on essay writing, it made Ornith slower.
MTP means multi-token prediction. A small draft head guesses the next few tokens, and the full model checks those guesses in one go.
▸ Guess right → you get extra tokens
▸ Guess wrong → that compute is wasted
▸ On JSON: Qwen 3.8 got 2.3× the speed with MTP, Ornith 35B got only 1.1×
▸ On essays: Ornith 35B dropped to 0.7×, so it became slower
▸ Why? On essays, the guesses are right only around 30% of the time
So test MTP on your own content first.
Will you test MTP? Tell me in the comments 👇
Follow for more AI benchmarks.
#MTP #LLM #AI #Shorts #Qwen #LocalLLM #AIBenchmark
Full benchmark: kgptalkie.com/tutorials/llm-benchmarking/ornith-1-5-9b-vs-35b-a3b-benchmark










