6x inference performance opt
Fable 5 just performed a 6.5x decode performance increase and 7.2x prefill performance increase in the span of a 55 minute goal thus far in my inferencing engine for Nemotron models, using up 8% of my 5-hour limits and 3% of my weekly limits.
And it's found more.
Users thank @_sholtodouglas for Fable 5's reported 6.5x decode gains, faster performance than GLM 5.2, and direct thought-to-code translation.
No Digg Deeper questions have been answered for this story yet.
Cool knight game
Fable 5 is a beast at ai game Dev!
Sure a few bugs here and there, but my goodness this is amazing.
its funny, and self-aware
crazy rl creature:
Using Fable 5 to add fur and a new voice to my RL creature. The voice is based on the neural network activation (pictured top left). Made with @threejs @webgl_webgpu @runpod trained with JAX and MuJoCo from @GoogleDeepMind
"GLM 5.2 was significantly more expensive, and wayyyy slower, than Fable 5"
This honestly surprised the heck out of me, GLM 5.2 was significantly more expensive, and wayyyy slower, than Fable 5 🤯
"I missed this model so much."
Fable 5 (max) “build a clash of clans clone”
Fable absolutely destroys Sonnet 5, and I’ll show that in an upcoming comparison video.
I did give it one follow-up while it was working, basically telling it I only needed a demo and didn’t need it to overthink the whole thing, so it ended up being only around 2,500 lines.
The funny part is it still grabbed the Clash of Clans assets with no trouble, which genuinely surprised me.
I missed this model so much.
"Fable is the first model where I've started to think of it as a coworker, and where I actually have trust in its system design."
Fable is the first model where I've started to think of it as a coworker, and where I actually have trust in its system design.
Gone are the days of excessive try-catch statements, slop, and rewriting the same helper function 10 times in the codebase.
Let's rip!
More inference optimisation
Fable is a monster. 4 hours max thinking goal mode.
It got Qwen 3.6 27B at 100+ TPS on a INT8/BF16 hybrid version with INT8 KV cache at 100k context window on 2x 3090s with 8 sessions
The model is 34GB for reference.
@elliotarledge cannot wait to see it on your kernel bench.
Amazing CFD sim:
It did a fantastic job making a web CFD simulation. It almost one shot this! I just had to ask it to add supersonic, and it did that too. You can access it here:
https://browertech.tplinkdns.com/physics/cfd-simulator.html
"Fable took a look and improved it 100x while providing -me- context I didn’t know I needed, and then doing a few quality of life things for me I didn’t ask for, but 100% love. Meaningfully and tangibly superior."
Same. And not just coding.
I built a client proposal using Opus 4.8 and GPT 5.5 Pro over the course of 5 days. I was very happy with it.
Then Fable took a look and improved it 100x while providing -me- context I didn’t know I needed, and then doing a few quality of life things for me I didn’t ask for, but 100% love.
Meaningfully and tangibly superior.
"it's the most direct 1:1 translation of thought-to-reality i've interfaced with so far"
the best way i can describe fable is a 100x increase for navigating my world of ADHD
my super power is i can spawn in ideas at whim and generate a design in my mind
with lesser models (5.5 was almost there but not quite) it takes longer to align their token output with my inner intent
with fable i can just beam this shit into my fucking computer and with 0 hand-holding, or correction it just brings my will to reality
it's the most direct 1:1 translation of thought-to-reality i've interfaced with so far
(this isn't even considering META's latest research on telepathic neural net transmitters LMFAO)
so if i can BEAM IDEAS without BABYSITTING them for correction, or long pointless discussion, because the model was just not smart (or big) enough
and then just progress. and move forward. literally move at the speed of thought (and compute it takes to processs that thought)
i consider it the greatest reality compression algorithm i have access to right now
no wonder i got depressed when it was taken away you gave me DIGITAL HEROIN and injected it STRAIGHT INTO MY VEINS and i wondered why i was tweaking !
what the fuck!
The race for superintelligence:
omg... Fable 5 is back and it's so good
I asked Fable to create a game called: The race for Super Intelligence...
After 4 prompts, and $173 worth of tokens Fable 5 created this game. (Prompt Below)
Rocket league:
"it's just machine gunning PRs."
holy. shit.
fable is an absolute monster. the output quality is wild, crisp, comprehensive, FAST, and the aesthetics are just gorgeous. it's just machine gunning PRs.
i don't mean to gush but i'm floored.
i don't need a benchmark chart telling me "akchually swe-bench pro is 2 points higher." if you can't viscerally feel a model's performance by now, no chart is going to help you.
@_sholtodouglas Definitely was surprising to me, and still more benchmarking to do. I really want to understand better why GLM 5.2 ends up using so many more tokens.
@_sholtodouglas From the man himself??
@morganlinton follows from this chart - for a given level of performance, smarter models need to use less tokens
@_sholtodouglas Thank you Mate :)