Compiler and optimization engine for artificial intelligence.
Bittensor
Joined April 2025
- Tweets815
- Following300
- Followers1.6K
- Likes6.4K
True Performance Network retweeted
TPN first consumer facing product will launch in November
True Performance Network retweeted
NEW: @TPN_Labs has added Quantization Aware Distillation to TPN Bench on Subnet 65. It retrains a quantized model against the original model to help recover behaviour lost in rounding.
xitter.cf/TPN_Labs/status/211354…
Quantization Aware Distillation available on TPN Bench.
Users can quantize their models with TPN Bench toolkit, and can choose to do so at 3 different levels: Standard, I-Matrix and Quantization Aware Distillation.
Here's how each work and what makes QAD a valuable tool for devs, labs and independent users.
Standard quantization:
The model’s parameters are stored at 16 bit precision. Quantizing rounds them down to 8 bit, or lower, like a zip file version of the model, making it smaller.
I-matrix:
Same idea, but it looks at the model first and finds the regions that appear most complex, assuming that complexity is where the capability lives. Those regions keep higher precision while the rest gets rounded harder.
Quantization Aware Distillation:
This method quantize the model, then treat the original model as a teacher and the quantized version as a student, training the student back toward its parent’s behaviour. It repairs what the rounding broke, needs inference and back propagation, meaning GPU work, which is why most people don’t do it, and why it’s the one worth having a service for.
Quantize LLMs with Bench
Quantization Aware Distillation available on TPN Bench.
Users can quantize their models with TPN Bench toolkit, and can choose to do so at 3 different levels: Standard, I-Matrix and Quantization Aware Distillation.
Here's how each work and what makes QAD a valuable tool for devs, labs and independent users.
Standard quantization:
The model’s parameters are stored at 16 bit precision. Quantizing rounds them down to 8 bit, or lower, like a zip file version of the model, making it smaller.
I-matrix:
Same idea, but it looks at the model first and finds the regions that appear most complex, assuming that complexity is where the capability lives. Those regions keep higher precision while the rest gets rounded harder.
Quantization Aware Distillation:
This method quantize the model, then treat the original model as a teacher and the quantized version as a student, training the student back toward its parent’s behaviour. It repairs what the rounding broke, needs inference and back propagation, meaning GPU work, which is why most people don’t do it, and why it’s the one worth having a service for.
Quantize LLMs with Bench
True Performance Network retweeted
Before renting a bigger GPU, I'd want to know whether the model could be smaller.
@TPN_Labs SN65 reports compressing a 17GB Mistral model to 7.71GB, with a higher score on its fluid_knowledge test: 74.93% versus 66.57%.
That's a useful direction for $TAO Bittensor. Make an existing model practical for more people to run, then compete on how much capability survives the compression.
A smaller file doesn't guarantee a smaller inference bill. I'd test memory use, speed and answer quality on the actual job. That's where this result could become a reason to choose the product.
TPN Labs is on @LinkedIn
Do you have a Linkedin profile too? Let’s connect!🤝
🔗 linkedin.com/company/trueper…
True Performance Network retweeted
I always prefer the Bittensor subnets that actually ship a consumer product.
Most of them stop at models and leaderboards. The ones that build something you can just open and use stand out to me.
TPN moving toward a consumer interface is the right direction. You can recommend it to anyone they don’t need to understand the technical setup. As long as it runs on their phone or desktop, that’s enough.
Real adoption starts when the product is usable anytime, not just when miners are competing.
bittensor:native
Upcoming: TPN Product Launch
In November, we’ll launch a consumer facing product powered by TPN.
The product will serve as one of the distribution layers for TPN Labs models, acting as an interface between our models and every day users.
We are excited for what's to come!
True Performance Network retweeted
NEW: @TPN_Labs has released TPN Model 006, a 7.71GB compressed model. The team reports a 74.93% fluid_knowledge score, versus 66.57% for the 17GB target model.
xitter.cf/TPN_Labs/status/211252…
TPN Model 006 is live on @huggingface
The @MistralAI target model (17GB) scored 66.57% on fluid_knowledge.
TPN-006, its compressed version, scored 74.93% at a size of 7.71GB.
006 also set the agenda for TPN-007, adding revised KL divergence and quantization logic, with a wider set of supported architectures.
In house teams can take weeks to successfully optimize a single model, TPN produced several iterations of the base model in a span of days, each shipped from individual teams working independently.
Find the HF repo to the winner's model below 👇
TPN Model 006 is live on @huggingface
The @MistralAI target model (17GB) scored 66.57% on fluid_knowledge.
TPN-006, its compressed version, scored 74.93% at a size of 7.71GB.
006 also set the agenda for TPN-007, adding revised KL divergence and quantization logic, with a wider set of supported architectures.
In house teams can take weeks to successfully optimize a single model, TPN produced several iterations of the base model in a span of days, each shipped from individual teams working independently.
Find the HF repo to the winner's model below 👇
New Quantization Service
TPN Bench now offers a full Model Quantization Toolkit, which includes methods like I-matrix and Quantization-Aware Distillation.
Here's how it works:
Users can upload any Hugging Face model and select the quantizations they want;
TPN Bench will generate those quantizations, upload them to a bucket, and then send back the links to those quants. It can also upload them to Hugging Face if given access to it.
Choose between regular quantization, I-matrix, or more advanced techniques like quantization-aware distillation, which require heavy infrastructure that we run for you.
Simplify the whole quantization model pipeline with TPN Bench!
🔗 bench.trueperformancenetwork…
True Performance Network retweeted
🚨 One week. 8 $TAO subnets. 8 different industries. Zero of them sell the same thing.
Oct 12 to 18, from @opentensor 's weekly highlights:
• SN92 @reaxlabs: open inference engine up to 4x faster on the same GPU, plus an MIT-licensed open-weights router (26 ms on a CPU)
• SN65 @TPN_Labs: Ministral-3-8B made 40% smaller, full 32K context kept
• SN95 @actualinc: 10,500 tok/s on a single NVIDIA DGX Station, launched with NVIDIA
• SN53 @engyai: 50B tokens in a day, then nearly 102B the next day on its public dashboard
• SN25 @urnetwork: 100,000+ existing providers moving onto mainnet as miners
• SN44 @webuildscore: computer vision now at a Shell motorway station, after the Avia rollout
• SN80 @openroboto: robot training data devices with on-chain identities on Solana
• SN28 @say_gm_: Claude, GPT and Gemini in one desktop app, priced below list
Decision models. Model compression. Inference hardware. Token serving. Bandwidth. Vision. Robotics data. A consumer app.
That's the whole AI stack, shipped in 7 days.
To match this week elsewhere you'd need:
• 8 companies
• 8 cap tables
• 8 budgets.
Big Tech does it with a payroll.
Bittensor does it with a market.
One token pays all 8 teams to compete, and the market decides who keeps getting paid.
Name one project in crypto that ships across the stack like this.
I'll wait.
Yes, true, its not perfect and most benchmarks are team-reported. Most Subnets are early and uneven.
But a week like this, every week, is how a market gets mispriced.
Don't trust me.
Check it yourself: engy.ai/usage, decisionmodels.io, and Actual's NVIDIA case study.
$TAO / DYOR.
True Performance Network retweeted
TPN Labs produced its first usable model
TPN Labs produced its first release-ready model.
The model (TPN005) is the outcome of an optimization process that reduced the size of the Ministral-3-8B base model by ~40%, while keeping its full 32k context.
This was the first round scored entirely on Fluid Bench, and it marks the first successful TPN upgrade.
So, what now?
We are happy with this result. This was Fluid Bench's first iteration, which paved the way for even better models to be produced by Subnet 65, with many more patches and upgrades to be implemented along the way.
Subnet produced models will become highly relevant once we release TPN's upcoming consumer facing product.
To close Competition 5 on a high note, we'll release a case study with all the numbers shortly.
True Performance Network retweeted
Very excited for this; this steam kettle been cookin' with the lid on for too long
Next month is going to be a big one for TPN 👀
We’ve been building something that we want to share with the world.
Powered by TPN_Labs - Subnet 65
$TAO
True Performance Network retweeted
November is going to be so much fun. I can barely believe the amount of stuff we're shipping.
Next month is going to be a big one for TPN 👀
We’ve been building something that we want to share with the world.
Powered by TPN_Labs - Subnet 65
$TAO