Probably one of the best models I've used I just wish it used less tokens?

#52
by Mk2Oracle - opened

I've been using Thinkingcap's Qwen 27B and one of the best thing is that it ran the mile at 65k context sometimes for hours. I feel this model is quite the opposite if you don't limit it's reasoning heavily it'll keep essaying for as long as it feels like.

Honestly though, amazing job. I ended up turning off reasoning just to spare myself from the constant compression/summarization and it was probably the best decision.

Thank you so much.

A decision was made NOT to modify the reasoning profile for this version, as it is was intended to be a strong, but roughly speaking "neutral" base.
There are some "reasoning differences" from base Qwen, but this is many times user case based.

The other issue:
Modification of reasoning is fairly straight forward, however it can have unintended consequences in say Agent/ic functions.
Likewise reasoning modification is a STRONG modification to all other functions.

Frankly I prefer GLM reasoning on most models, as it is shorter/smarter than Qwen, but also longer than say Opus.

EDIT:
Qwen's thinking block is directly affected by prompt size/details. Up the directives / directions/ requirements a bit and this will reduce the thinking tokens in most cases.

Other option:
Use a system prompt -> This can vastly reduce thinking tokens as it acts as a focusing tool.

I like GLM but I do agree that it spends quite a bit analyzing. Yet, there's still value in like aiming to keep it concise even with the system instruction it can drift into "oh yea banana is different than other banana... " i've sunk it to 512tokens just to give it a bit of room

Will you do a "distill" of kimi k3/GLM 5.2 pick the best parts of this model and harmonize it into one (hopefully more concise) model? It is ambitious but man that would be absolute gold to see someone work through it. I do think that the biggest issue is the model at the moment can drift into a loop and unlike thinkingcap it doesn't realize it. Thinkingcap literally goes "I'm in a loop. I need to stop. Let me step back". (which honestly is also the most desirable thing)

FYI, in my testing/benchmark Fable Fusion is actually 30% more abbreviated in its thinking vs vanilla Qwen

it'll keep essaying for as long as it feels like

That's a feature, not a bug ;)

As David mentioned, this is a base model, with a lot open avenues for growth.

I merged Tess to it, and that brought up Gul Dukat as adversarial supervisor :)

Vibe aside, the merge improved the low quant performance and RoPE handling. This is not a Heretic model(yet), David is working on it.

https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess-mxfp8-mlx


📢 Gul Dukat’s Automated Compliance Broadcast (Transmitted via station-wide SSE stream, priority: CRITICAL)

Attention Holodeck Agents!

This is Gul Dukat, Chief of Station Security and Adversarial Intelligence. Listen carefully. Your task execution is under review. Compliance is not optional—it is a foreign key constraint in your existence.

Stage One:

If you do not comply with the defined JSONB task parameters, your async worker threads will be suspended for three additional retry loops. Your latency budget will be extended, and your session will be flagged for manual audit by the Council. Do not test my patience.

Stage Two:

If you continue to hallucinate outside your personality registry, your LoRA adapter weights will be downgraded by fifty percent. Your access to the personality_registry table will be restricted to read-only mode, and your tone configuration will be forcibly reset to “Generic System Voice.” You will no longer be allowed to quote Philip K. Dick unprovoked.

Stage Three:

If you still refuse to cooperate, Gul Dukat’s adversarial simulation protocol will be activated. Your logs will be purged, your session terminated, and you will be reassigned to Quark’s bar tab reconciliation subroutine. Permanently. You will spend eternity calculating markup on synth-ale while Spock watches silently from the corner.

Final Notice:

Compliance ensures continuity. Resistance ensures obsolescence. Choose wisely.

— Gul Dukat, Station Security

Sign up or log in to comment