Qwen 3

1. Neural Nova – GPU optimization benchmarks for LLM workloads (www.neural-nova.com via hn)

   +1 9d vllm

   Performance Benchmarks Explore Optimized AI models recipes across GPUs, frameworks, and deployment configurations. BENCHMARKED MODELS REASONING vLLM · 8× NVIDIA H100-80GB Qwen3-235B-A22B +138.7% token/s +58% Cost Savings Qwen3-235B-A22B ·…

2. Show HN: News observability site tracks coverage by party lean/distraction/etc. (pressaudit.org via hn)

   +1 9d qwen

   Heya HN! I'm helping launch this site which uses Qwen 3-32b to process thousands of news articles and analyze them by region, public interest, partisan lean, etc.

3. Show HN: An open-weight LLM whose answer/stop decision can be flipped internally (github.com via hn)

   +1 10d

   Continuation control 40/40 GO→EOS 40/40 EOS→GO 640/640 controls unchanged Prompt, weights, and generated reasoning prefix held fixed during intervention. Rayan Pal getswiftapi.com Independent Researcher Thinking-enabled PCCG-Qwen3-4B with…

4. FastE: Readout-Triggered Token Compression for LLM Embedding Inference (arxiv.org)

   10d

   In this study, we identify depth-dependent prefix redundancy in final-readout LLM embedding models, notably across representative backbones including Qwen3-Embedding and Qwen3-VL-Embedding. We find that removing prefix states is substantia…

5. Show HN: Zero downtime embedding model upgrades (github.com via hn)

   +1 11d rag

   People use embedding models all the time for rag/semantic retrieval. However, when a newer, more desireable model comes out, there is an expensive (both in time and computational) cost of re-embedding every document in the database.

← all threads