AI Briefing#benchmark
The latest AI and developer news about #benchmark, with the original source and a short summary.
The latest AI and developer news about #benchmark, with the original source and a short summary.
Llama.cpp token speed measurements on the Zenbook A16 with Snapdragon X2 Elite Extreme.
Ran Qwen3.6 35B 4bit for multi-user use via vLLM Docker on 2x RTX 3090.
CWT reduced compute by **about 45%** while maintaining a **PPL difference of 1.7%** versus the baseline.
AI investment and performance have surged, but employment and public perception remain mixed.
Introducing a pure SQL queue that runs on Postgres without bloat.
It compares the token usage difference between Opus 4.6 and 4.7 using anonymous inputs submitted by the community.
On DGX Spark, Qwen3.6 performs similarly to Qwen3.5, but MTP significantly increases throughput.
LIDARLearn, a unified PyTorch library for 3D point clouds, has been open-sourced.
Boosted Qwen3.6-35B-A3B speed by 54% on an RTX 5070 Ti using --n-cpu-moe.
A benchmark, safety, and tensor analysis comparing Qwen 3/3.5 with HauhauCS, Heretic, and Huihui was released.
On the RTX 5060 Ti 16GB, Qwen3.6-35B-A3B ran about 10% faster than the prebuilt version.
Claude 4.7 produces 1.3–1.45x more tokens while showing a modest improvement in instruction compliance.
Combining an RX 9070 and RTX 5070 Ti, the author ran Qwen 35B Q4 at over 100tps.
Claude Opus 4.7's new tokenizer was measured to use more tokens than officially stated.
The Intel Arc Pro B70 was compared with NVIDIA RTX and AMD Radeon AI PRO on Linux.
In an apples-to-apples FP8 comparison, the advantage between Ironwood and B200 reversed depending on the model.
On A100 40GB, MIG significantly lowered protected-tenant p95 compared to shared vLLM.
They trained a fast and accurate 6-language OCR model on 12.2 million synthetic data images.
On 28 Zod tasks, Opus 4.7 had the same pass rate as 4.6 but was faster and more efficient.
On dual RTX 5060 Ti GPUs, Qwen 3.6-35B-A3B hit 21.7 tok/s at 90K context.
SwiftLM presented results of running a 35B Qwen model on a 24GB Mac with 40K~100K context.
In the Qwen3.6-35B-A3B GGUF benchmark, Unsloth quants claim the best size efficiency relative to KLD.