Research desk

Reproducible findings, not hot takes

Technical reporting for reverse engineering, cloud security labs, vulnerability analysis, AI tooling, and defensive testing.

Research articles: 216
On this page: 25

Data graphic: dot plot of word error rate lower is better from NVIDIA's own model card, Nemotron 3.5 Transcribe against its closest rival on three separate tracks: 4.71 vs 4.94 for Qwen3-ASR-1.7B in English a near tie, 0.23 points , 4.62 vs 6.49 for Qwen3-ASR-1.7B in multilingual 1.87 points , and 10.98 vs 13.37 for Parakeet-TDT-0.6B-v3 in long-form 2.39 points . A callout notes the BF16 weights are 67.5 GB.

watchResearch Nvidia Nemotron 3 5

Nemotron 3.5 Transcribe Posts 4.71% English WER in Open Weights, and Fits a 32 GB GPU Only in 4-Bit

NVIDIA's open-weight speech model beats Qwen3-ASR and Parakeet on its own card, but BF16 weights are 67.5 GB. Which formats fit 16, 24 and 32 GB cards.

Read the research

Data graphic: LFM3-1.7B vs Qwen3.5-9B. By Liquid AI's own figures, the 1.7B model wins three of six benchmarks against a roughly five-times-larger model, but trails by 17 points on GPQA Diamond 63.37 vs 80.40 and on RULER 128K long context 65.56 vs 87.11 .

watchResearch Liquid Ai

LFM3-1.7B vs Qwen3.5-9B: The 1.7B Model Wins on Tool Calls, Then Loses on Reasoning and Long Context

Liquid AI's own table shows LFM3-1.7B ahead on IFEval, BFCL and LiveCodeBench but 17 points behind on GPQA. Ollama 0.40.13 runs it on every desktop.

Data graphic: timeline of three Ollama releases. v0.40.11 on 24 Oct added speech on macOS and d3 models, v0.40.12 on 27 Oct added LFM3 and d1 on Apple and DFlash2, and the latest, v0.40.13 on 28 Oct, brought LFM3 to Linux and Windows, three days and then one day apart. Speech uses phonon, a 31M-parameter English model, and LFM3 is a 1.7B text model with a 256K-token context.

stableResearch Ollama

Ollama 0.40.11 to 0.40.13: An OpenAI-Compatible Speech Endpoint, d3 Decision Models, LFM3 and DFlash2

Ollama releases between 24 and 28 October add macOS speech over /v1/audio/speech, d3 decision models, LFM3 on Linux and Windows and DFlash2 drafts on MLX.

Data graphic: vendor-run scores for Nanbeige4.3-33B and Qwen3.8-27B on Terminal-Bench 2.1, DeepSWE, SWE-Bench Pro and GDPval, showing Nanbeige ahead on the first two and Qwen ahead on the last two.

watchResearch Nanbeige 4 3 33b

Nanbeige4.3-33B vs Qwen3.8-27B: Where the Vendor Benchmarks Diverge, and Which GGUF Fits 16, 24 or 32 GB

Nanbeige's new 33B open model leads Qwen3.8-27B on terminal and agent tasks in vendor-run tests, loses SWE-Bench Pro, and needs care to fit 16-32 GB.

Data graphic: Claude Code compacts conversations of 200K tokens or more after about 54 minutes idle, before the one-hour prompt cache expires

watchResearch Claude Code

Claude Code Now Compacts Idle Sessions Before the Cache Expires: What Changes for Pro and Max Limits

Claude Code now compacts idle 200K-token sessions before the prompt cache expires and warns Pro and Max users before costly resumes. Here is what changed.

More from the desk

Full archive →

216 research articles · showing the newest 25Research archive →