Qwen3.8-27B: The Thinking Dial, Measured
Part one of this series covered quality, quantization, and the 1M-context price tag. This one is the part that matters most if you’re running the model as an agent: the reasoning dial.
Veladan Programming Blog
...for all the programming issues I found.
Latest posts
Qwen3.8-27B-FP8: First Benchmarks on RTX PRO 6000 Blackwell
Qwen3.8 is out, and it’s a dense 27B that Qwen positions as the most capable generation in its open-model family. Before I commit my RTX PRO 6000 Blackwell to it, I’m running a four-part test: quality (this post), quantization (FP8 vs. NVFP4), long context (the 1M token claim), and thinking depth (the four reasoning levels). This is the first data point: the FP8 variant on BenchLocal.app, run…
The Lie of Open Source AI: From Open Weights to Full Transparency
In the current Artificial Intelligence ecosystem, the label “Open Source” has become an elastic, almost promotional term. As a heavy user of models like Gemma-4 and Qwen, I’ve realized there is an invisible but profound gap between the software we can download and the software we can actually understand.
A Minimal Method for AI-Assisted Software Development
There are two dominant ways people use AI to develop software right now. Both have obvious appeal. Both are traps — not because the ideas are wrong, but because they solve the wrong problem.
The Permission Trap: From Right of Action to Administrative Concession
A recent debate regarding forest fires in Spain sparked a philosophical realization about the nature of freedom in the modern state. The discussion started with a simple question: Is it forbidden to clear dead wood and leaf litter from forests to prevent fires?
Comparing LLMs on the HermesAgent-20 Benchmark
I’ve been testing several high-performance models on the HermesAgent-20 benchmark suite via BenchLocal.app. This specific benchmark is designed to evaluate the “agentic” qualities of a model: its ability to follow complex multi-step instructions, maintain state, and orchestrate tool calls effectively within a simulated agent environment.
Benchmarking poolside/Laguna-S-2.1-NVFP4 on RTX PRO 6000 Blackwell
Today I ran a series of local benchmarks comparing poolside/Laguna-S-2.1-NVFP4 against Qwen3.6-27B-FP8 on my RTX PRO 6000 Blackwell. At first glance, Laguna underperforms — but the HuggingFace model card claims better numbers than Qwen3.6 on SWE-bench and Terminal-Bench. So either my setup is suboptimal, or there’s a gap between the official benchmarks and real-world usage. Here are the raw…
The Inverted Cost Model: Why Your Orchestrator Should Be the Cheapest Agent
I had my multi-agent cost strategy backwards. The orchestrator is your heaviest token consumer, so it should be the cheapest model — not the smartest one.
Why FP8: Choosing the Right Qwen3.6 Quant for Agentic Coding
I replaced my local model in OpenCode from Qwen3.6-27B to Qwen3.6-35B-A3B in FP8. Here is why, and why the other two options — the smaller 27B and the NVIDIA NVFP4 variant — lost.
Pipeline vs S2S: Architecting for Control in Voice Local AI
Why the Modular Pipeline remains the king of production local Voicebots over native Speech-to-Speech.