Meta released Muse Spark Open on 15 October.
Six days later, it's live in Cacheon's optimization arena.
Can miners find the kind of inference efficiency gains they've found on DeepSeek V4.1-Flash, GLM-5.3, and Qwen 3B?
Let's find out.
Six days later, it's live in Cacheon's optimization arena.
Can miners find the kind of inference efficiency gains they've found on DeepSeek V4.1-Flash, GLM-5.3, and Qwen 3B?
Let's find out.
0
1
7
879
DeepSeek has open-sourced DeepInfer, its optimized inference engine.
With performance benchmarks showing great margins for inference providers of close to 80%.
Cacheon, however, remains ahead of the curve.
Our latest DeepSeek-V4.1-Flash benchmark: 434 tokens/s per user, serving 24 users concurrently.
With performance benchmarks showing great margins for inference providers of close to 80%.
Cacheon, however, remains ahead of the curve.
Our latest DeepSeek-V4.1-Flash benchmark: 434 tokens/s per user, serving 24 users concurrently.
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
0
7
11
613
We're excited to be providing inference on the Liquid Inference exchange.
Token volume is picking up, so this presents a great opportunity to monetize our optimized catalog of models.
Prices are currently competitive but we expect to be out-competing and delivering on revenue.
Token volume is picking up, so this presents a great opportunity to monetize our optimized catalog of models.
Prices are currently competitive but we expect to be out-competing and delivering on revenue.
The Liquid Inference router network is growing quickly. We now have 28 verified companies serving >1000 models, offering capacity dynamically and updating quotes in real-time. Average time for providers to onboard and integrate models is 3 hours, full verification within a day.
0
10
23
1,865
More throughput. More ways to compete.
A 50% throughput gain on the same hardware, at equivalent quality, opens up new opportunities for teams building and serving AI:
→ Lower costs: offer more competitive prices, or make your own model cheaper to run and use
→ Compete on capacity: serve more demand before adding infrastructure.
→ Protect margins: stay viable as competitors reduce prices.
→ Make more workloads economical: bring previously expensive use cases within reach.
Whether your advantage is the model itself or how efficiently you serve it, better inference economics matter.
A 50% throughput gain on the same hardware, at equivalent quality, opens up new opportunities for teams building and serving AI:
→ Lower costs: offer more competitive prices, or make your own model cheaper to run and use
→ Compete on capacity: serve more demand before adding infrastructure.
→ Protect margins: stay viable as competitors reduce prices.
→ Make more workloads economical: bring previously expensive use cases within reach.
Whether your advantage is the model itself or how efficiently you serve it, better inference economics matter.
0
2
7
292
Why DeepSeek-V4.1-Flash for our next competition?
Because targeting a frontier open model keeps Cacheon at the leading edge.
Strong coding and agent performance. Native visual understanding. An architecture designed for efficiency.
How much faster can we run the same hardware?
Because targeting a frontier open model keeps Cacheon at the leading edge.
Strong coding and agent performance. Native visual understanding. An architecture designed for efficiency.
How much faster can we run the same hardware?
0
5
15
896
Affine's King model. Two weeks of optimization. 40.9% more throughput.
@affine_io became Cacheon’s first customer. These are the performance numbers behind that partnership.
We pushed Qwen3.6-35B to 528.9 tokens/s, up from 375.4 tokens/s on stock SGLang.
We’re building on this n’s inference optimization to more customers and models.
@affine_io became Cacheon’s first customer. These are the performance numbers behind that partnership.
We pushed Qwen3.6-35B to 528.9 tokens/s, up from 375.4 tokens/s on stock SGLang.
We’re building on this n’s inference optimization to more customers and models.
2
12
25
2,070
$100k in revenue last month. @affine_io became our first customer.
We’re now working to monetize our Qwen and GLM-5.3 optimizations through licensing deals and our own inference service.
This quarter, we’re targeting more customers and expanding competitions to high-volume models - building a catalogue of kernels we can license or deploy ourselves.
There's more to come.
Keep calm and Cacheon.
We’re now working to monetize our Qwen and GLM-5.3 optimizations through licensing deals and our own inference service.
This quarter, we’re targeting more customers and expanding competitions to high-volume models - building a catalogue of kernels we can license or deploy ourselves.
There's more to come.
Keep calm and Cacheon.
0
6
35
1,931
Q3 is over. How are we doing against our roadmap?
The proof of concept is built:
✓ Multiple models optimized - MiniMax-M3., GLM-5.3 and Qwen3.6.
✓ Competitions run in parallel across models.
✓ Speed improvements evaluated end to end, with output-quality checks.
Different obs. This progress opens a path to serving more workloads, with improvements judged on both speed and quality.
Q4’s focus: monetization and product–market fit.
The proof of concept is built:
✓ Multiple models optimized - MiniMax-M3., GLM-5.3 and Qwen3.6.
✓ Competitions run in parallel across models.
✓ Speed improvements evaluated end to end, with output-quality checks.
Different obs. This progress opens a path to serving more workloads, with improvements judged on both speed and quality.
Q4’s focus: monetization and product–market fit.
0
4
8
552
AI inference is getting cheaper. But it’s consuming more and more of the AI compute stack.
McKinsey released a new report yesterday, estimating that by 2030, inference will account for ~60% of AI compute demand, versus ~40% for training.
That makes inference optimization increasingly consequential: every gain in throughput and utilization changes how much useful AI the same hardware can serve.
McKinsey’s full report ↓
mckinsey.com/industries/sem…
McKinsey released a new report yesterday, estimating that by 2030, inference will account for ~60% of AI compute demand, versus ~40% for training.
That makes inference optimization increasingly consequential: every gain in throughput and utilization changes how much useful AI the same hardware can serve.
McKinsey’s full report ↓
mckinsey.com/industries/sem…
0
7
17
1,523
Great to see the work we’ve been doing with @affine_io get a moment on stage at @ExploitSummit.
Here’s @const_reborn telling the story in his opening keynote ↓
Here’s @const_reborn telling the story in his opening keynote ↓
0
5
19
1,963
Qwen3.6 arena is now open for submissions!
Better yet, we've generalized the evaluation system to support arbitrary kernel changes.
No more waiting on slot updates to submit your winning kernel!
We're excited to see what you can do. Submit today and earn Alpha right away.
Better yet, we've generalized the evaluation system to support arbitrary kernel changes.
No more waiting on slot updates to submit your winning kernel!
We're excited to see what you can do. Submit today and earn Alpha right away.
0
3
9
782
GLM-5.3 hits a new milestone on Cacheon: ~12.5% above its earlier benchmark, measured under different test conditions.
On SN14 in $TAO Bittensor, miners optimize GPU operations to accelerate inference. We test changes with the full model running, and accepted work can support Our latest competition submission reached 1,242.8 tokens per second. We're contributing to raise the bar and challenge other miners to push GLM-5.3 further. Our team follows the same evaluation, submission-order and acceptance rules as every other miner. We'll continue contributing as long as our involvement helps advance the subnet. Any emissions we earn will fund development and operations.
In parallel, we're working to turn these advances into commercial opportunities, including collaborations with other subnets. We'll share more soon!
On SN14 in $TAO Bittensor, miners optimize GPU operations to accelerate inference. We test changes with the full model running, and accepted work can support Our latest competition submission reached 1,242.8 tokens per second. We're contributing to raise the bar and challenge other miners to push GLM-5.3 further. Our team follows the same evaluation, submission-order and acceptance rules as every other miner. We'll continue contributing as long as our involvement helps advance the subnet. Any emissions we earn will fund development and operations.
In parallel, we're working to turn these advances into commercial opportunities, including collaborations with other subnets. We'll share more soon!
0
10
22
2,259
In our last competition, Cacheon miners pushed MiniMax M3 throughput to 2,337.9 tok/s: +32.2% vs the SGLang baseline.
What might that kind of performance improvement be worth commercially?
+32.2% throughput = ~61% more profit.
How? SemiAnalysis InferenceX™ models a B300 LLM at $4.81 profit per chip-hour. Apply the same +32.2% throughput gain, with pricing and chip costs held constant, and modeled profit rises to $7.73 per chip-hour.
Why the difference? The GPU still costs $4.25 for the hour. More throughput means more capacity to sell without buying another chip-hour.
That's a scenario, not measured Cacheon profit. What is measured is the optimization: 2,337.9 tok/s, +32.2%, produced by Bittensor miners and passed through Cacheon's testing and reproduction.
Check the baseline. Run your own numbers.
What might that kind of performance improvement be worth commercially?
+32.2% throughput = ~61% more profit.
How? SemiAnalysis InferenceX™ models a B300 LLM at $4.81 profit per chip-hour. Apply the same +32.2% throughput gain, with pricing and chip costs held constant, and modeled profit rises to $7.73 per chip-hour.
Why the difference? The GPU still costs $4.25 for the hour. More throughput means more capacity to sell without buying another chip-hour.
That's a scenario, not measured Cacheon profit. What is measured is the optimization: 2,337.9 tok/s, +32.2%, produced by Bittensor miners and passed through Cacheon's testing and reproduction.
Check the baseline. Run your own numbers.
0
9
20
1,112
Happy Labour Day!
We've been hard at work over the weekend.
Announcing our GLM-5.3 arena!
Miners can now submit kernel optimizations and get rewarded by the protocol.
GLM-5.3 presents a huge opportunity for production-use and monetization of the subnet.
More info: cacheon.ai
We've been hard at work over the weekend.
Announcing our GLM-5.3 arena!
Miners can now submit kernel optimizations and get rewarded by the protocol.
GLM-5.3 presents a huge opportunity for production-use and monetization of the subnet.
More info: cacheon.ai
0
4
11
1,275
Another great result coming out of this competition's arena.
We've hit 2,274 tok/s on our MiniMax-M3 baseline
We're hard at work on integrating the next top model for our arena.
Cacheon focuses on production-ready models for real enterprise use-cases with evaluations only via re excited for the submissions we'll get on our next arena. More to come.
We've hit 2,274 tok/s on our MiniMax-M3 baseline
We're hard at work on integrating the next top model for our arena.
Cacheon focuses on production-ready models for real enterprise use-cases with evaluations only via re excited for the submissions we'll get on our next arena. More to come.
0
8
32
3,950