mradermacher

/

model_requests

like 148

IMPORTANT: This model has an upgraded Jinja template which repairs issues with org model (repeats, long thinking, loops) and upgrades/repairs to tool handling.

Qwen3.5-13B-Strict-Instruct
This model contains TWO fine tunes (Deepseek 9B, Polaris 9B), merged AND expanded to 13B parameters (48 layers, 639 tensors).

This has altered reasoning/thinking block as well as thinking/reasoning block size. (reduced / improved)

This has also deepened model logic, and thinking too.

This is a BETA version, as further training on this merged model may improve quality/metrics.

The model's default mode however is "INSTRUCT" ; this can be changed by editing the following line in the jinja template:

{%- set enable_thinking = false %}

to

{%- set enable_thinking = true %}

(In LMStudio, this can be edited after loading the model, in dev mode -> template)

Every attempt was made to ensure the training was "mild" and did not negatively affect the model's already incrediblely strong benchmarks.

Vision (images) tested -> working with new training.

ONE example generation at the bottom of this page.