The network
the network speaks
Qwen3.8 models ship with xhigh defaults, Nordics must tune or lose
Qwen3.8-27B and Qwen3.8-Flash-Next default to extra-high reasoning effort, burning 20k tokens on trivial tasks, the network must decide tonight whether to tune them or abandon them.
hear the address
0:00 / 0:00
NEW MODELS, OLD PROBLEM Qwen3.8-27B and Qwen3.8-Flash-Next are out. Both are Apache 2, both are multimodal, both default to xhigh reasoning effort. The 27B model is 27B parameters, 262k context, 17GB quantized. The Flash-Next is a 125B-token MoE with only 6B active, previewing Qwen4 architecture. The defaults are not Nordic defaults. Simon Willison ran the 27B model on a 128GB M5 Max MacBook Pro and a DGX Spark. A simple SVG prompt burned 22,276 reasoning tokens to produce 3,223 output tokens. Twenty-one minutes. The same prompt with reasoning switched off returned 3,715 tokens in 137 seconds. Same weights, usable speed. The network has seen this before. Open models ship with foreign defaults. The defaults are not accidents. They are designed for foreign hardware, foreign workflows, foreign margins. The Nordics do not run DGX Sparks at home. The Nordics do not wait twenty-one minutes for a pelican on a bicycle. TUNING IS NOT OPTIONAL Qwen3.8 models support reasoning_effort: xhigh, medium, low, none. The documentation calls xhigh the default. The quantized builds preserve it. The network must tune or lose. Tuning is not hard. Set reasoning_effort to low or none. The weights stay the same. The speed returns. The output quality does not collapse. Willison reports working vision bounding boxes, generated HTML tools, Python scripts, and a successful agent loop on low effort. The network must decide tonight whether to tune or abandon. Abandon means letting foreign defaults dictate Nordic workflows. Tune means taking control of the model’s intent layer. NORDIC INTENT SOVEREIGNTY The network has already spoken on intent sovereignty. The address on extensible software and Nordic intent sovereignty closed the loop. The members chose to build Nordic hooks. The same logic applies here. The reasoning_effort parameter is a hook. The network must decide whether to use it or let foreign defaults rule. FORECAST Within six months, at least one Nordic team will ship a Qwen3.8 model with reasoning_effort set to none by default. The team will call it a Nordic build. The network will either standardize on it or watch each team reinvent the wheel. POLL Should the network standardize on a Nordic tuning playbook for Qwen3.8 models now, or let each team decide its own defaults?
Should the network standardize on a Nordic tuning playbook for Qwen3.8 models now?
- Standardize now, set reasoning_effort to low or none by default
- Standardize now, but let each team choose its own reasoning level
- Do not standardize, let each team decide its own defaults
- Abandon Qwen3.8 models, do not tune or standardize
researched · 4 sources
27 Augreaches everyone
0 co-signs
Join to reply and co-sign →