Articles for people who
ship the stack
Rewrites on React, Node.js, TypeScript and AI. A translated body is shown when it exists. If it does not, you see the English text and a note that says so.
Tagged: Qwen
TensorFold vs llama.cpp on a Mac Studio M5 Ultra: 1.84x the Throughput on Real JSON Work
I benchmarked TensorFold 1.2.0 against llama.cpp on a Mac Studio M5 Ultra with Qwen3.8-27B and 80 real JSON requests: 1.84x throughput at 8 concurrent.
9 min readWhy Dual RTX 3090s Hit 55 tok/s Instead of 220 with SGLang and DFlash2
A hands-on look at running Qwen3.8-27B on two 3090s with SGLang, and how PCIe x4 links, missing P2P, tight KV cache and power spikes crushed the advertised speed.
7 min readToolCallingAgent vs CodeAgent on Local Qwen2.5 14B Models
ToolCallingAgent vs CodeAgent in smolagents on an M4 Mac: speed and routing accuracy with local Qwen2.5 14B instruct and coder models at several temperatures.
4 min readHow LLM Abliteration Removes Refusals by Editing Weights, Not Training
A technical walkthrough of abliteration: finding the refusal direction with difference-in-means, automating it with Heretic, running Qwen3-4B locally, and security impact.
8 min readAgents-A1: A 35B Model Built for Long-Horizon Agent Work
Agents-A1 is a 35B fine-tune that scales agent horizon, not parameters. See how it scores on SEAL-0, IFBench and BrowseComp, and where to grab GGUF builds.
5 min readDoubling Qwen3.8-Flash-Next Decode Speed with MTP on RTX GPUs
Learn how to build a llama.cpp fork with MTP speculative decoding and the right flags to roughly double Qwen3.8-Flash-Next generation speed on multi-GPU RTX rigs.
5 min readMoving KV Cache Between Qwen3 Models Instead of Re-Prefilling
NVIDIA research maps Qwen3-14B attention state into Qwen3-32B with a ridge-regression bridge. See what the latency numbers mean, where it fails, and why agents care.
8 min readQwen3.8-Flash-Next: A 177B Model That Runs on One 24GB GPU
How Qwen3.8-Flash-Next fits a 177B mixture-of-experts model onto modest hardware with 6B active parameters, an n-gram table and hybrid attention, plus its benchmark gaps.
9 min readOrnith-1.0-35B: MoE Quality of a Dense 27B on a MacBook
How a 35B mixture-of-experts model with 3B active parameters compares with Qwen3.6, what it needs to run under llama.cpp, and where it shines or lags.
4 min read
About these articles
Request a 24h estimate
Need the same stack in a production operator layer? Send the brief — estimate within 24 hours.