Microtensor | Sn92

Certified inference systems, measured on the hardware they claim | Bittensor SN92

[Microtensor | Sn92 @Microtensor92]

Arena 6 is live: hallucination detection

The sixth microtensor competition opens on subnet 92: guard / mt-4g.

This arena transitions our focus is high demand enterprise utilities, specifically targeting general hallucination detection to guarantee logical consistency, accuracy and data security in LLMs.

The task: Hallucination detection. Identifying factual contradictions, data anomalies and logic lapses in model outputs.

The envelope:

disk 3.2 GB
memory 4.3 GB
p95 latency 45,000 ms

Base models:
`google/gemma-3-4b-it`
`meta-llama/Llama-3.2-3B-Instruct`
`microsoft/Phi-4-mini-instruct`

The dataset is attached once submissions begin.

Arena details: https://t.co/ijiQqlg32b

[Microtensor | Sn92 @Microtensor92]

Microtensor is building a compute marketplace where miner-supplied hardware can move between workloads and serve multiple forms of demand:

→ Certified inference
→ Bittensor subnet mining
→ GPU rentals
→ Model shaping

When inference demand rises, hardware serves inference.

When capacity is available, it can mine.

When someone needs dedicated compute, it can be rented.

The objective isn't to create another pool of idle GPUs, but to make every unit of compute productive.

And as the network expands beyond mining, we're aligning miner participation with the people building for the long term.

Microtensor Compute is coming.

[Microtensor | Sn92 @Microtensor92]

Microtensor’s official community call is coming up.

We’ll be walking through the next phase of SN92: Microtensor Compute, how the compute marketplace works, how miner-supplied hardware can serve different workloads, and how the incentive mechanism is designed to create more value across the network.

We’ll also introduce the new miner commitment mechanism, what it means for submissions, and why we’re putting greater emphasis on miners who are committed to building with the subnet for the long term.

We’ll cover what’s changing, why we’re making these changes, and what miners, validators, and the wider Microtensor community should expect going forward.

This call is public. Miners, validators, builders, stakers, and anyone interested in SN92 are welcome.

12th September// 3pm EST // https://t.co/taDRvHg60W

[Microtensor | Sn92 @Microtensor92]

Round 1,240 has settled. Arena 5 (`support / mt-3g`) is complete.

49 submissions came in. 43 systems were frozen and measured, 42 were admitted, and 6 never reached measurement and settled unscored. Validators produced 121 signed reports at replication 2 with a divergence rate of 0.0 between paired measurements.

Rewarded, by exclusive hypervolume

# uid hotkey quality cost (p95 ms) share
1 42 5CLfzJdv... 0.9556 128 0.30
2 39 5CUWFXGB... 0.9722 4461 0.20
3 47 5DoaV8JK... 0.9556 457 0.14
4 26 5EyYXepK... 0.9556 803 0.11
5 9 5CcaCist... 0.9522 1285 0.09
6 23 5FnxoF33... 0.8018 457 0.07
7 34 5HEtPjha... 0.9087 2104 0.05
8 209 5CCaDSju... 0.9083 2225 0.04

On the reserve: Shares above are of the miner half. The reserve holds 50% this round and remains temporary. Compute testing is now open, and once it is live the emission share rebalances and the reserve comes down.

Note on cost: This was the first round scoring on p95 total task latency rather than time to first token, so an empty answer could no longer buy a cheap position.

The paid frontier now spans 128 ms to 4.5 s. The highest quality system on the board (0.9722) sits at 4.5 s and takes the second largest share, while three systems tied at 0.9556 are separated purely by cost and take first, third and fourth. The fastest system on the board, at 128 ms, took the largest share.

All rewarded systems are archived with their model weights, manifests and certificates under the base model's licence (CC BY-NC 4.0): https://t.co/heWMJAb3XU…

Leaderboard: https://t.co/sxeH5SrOIe…

[Microtensor | Sn92 @Microtensor92]

Most AI networks are focused on models.

Microtensor is focused on the systems running those models.

The interesting part is that those systems can be measured on quality, cost, latency and energy, then certified and made reproducible.

That’s a much more interesting problem to solve.

[Microtensor | Sn92 @Microtensor92]

Microtensor Compute testers wanted

We're opening testing for Microtensor Compute, and we want miners involved early.

What it's for: The network produces certified systems, and the registry lets anyone download and run them. Compute is the layer that serves them: inference for certified models, hosted on the network rather than a single provider. A developer who doesn't want to run weights themselves gets an endpoint, and the compute behind it comes from the network.

That also closes the loop on emissions. Compute is what the reserve is funding, and once it's live the emission share rebalances and the reserve comes down.

What your GPU will do: Four things, and the pool decides which at any moment rather than you picking a lane.

• Inference on certified systems, through a routed API where a customer states a quality floor and gets the cheapest certified system that clears it
• Rental, as clusters, sandboxes and storage
• Model shaping, meaning fine tuning, custom training and certified evaluations for people who bring their own model
• Subnet mining with whatever is left over, so a rig is never idle

This is a testnet: Things will break, and telling us how is the point. Post failures with the classification and seed from the validator log.

Setup and requirements: https://t.co/za63jL4NTz

Join here: https://t.co/C0UOdKL3m5

[Verathos @verathos_ai]

Verathos x Microtensor

We’re partnering with @Microtensor92 to bring certified subnet native intelligence to Verathos. As a first step, we’ve deployed their certified coding model, mt-code-3g-r1238-5ED3LzEW, to the Verathos testnet.

The model comes from Microtensor’s round 1,238 arena, built on Qwen2.5-Coder and licensed under Apache 2.0. Its performance credentials are backed by a verifiable certificate, making it a strong fit for Verathos’ focus on accessible, verifiable AI.

This integration expands what Sleipnir can route to. Instead of serving only general purpose models, Verathos can increasingly make specialized intelligence produced across Bittensor subnets accessible through shared infrastructure.

It’s an early step toward making the intelligence created across Bittensor accessible beyond the subnet where it originated.

Explore the model: https://t.co/QG9RP2sbjw
Try it on Verathos Chat: https://t.co/1ylbZp4Ods

Switch to Testnet in the top right selector.

[Microtensor | Sn92 @Microtensor92]

@verathos_ai *

[Microtensor | Sn92 @Microtensor92]

Microtensor × Verathos

Certified inference is starting to travel across the Bittensor ecosystem.

@Verathos has deployed "mt-code-3g-r1238-5ED3LzEW" from the Microtensor archive to their testnet, making the certified system available through their inference network.

The model was measured by Microtensor validators and is now being served beyond the subnet that produced it.

Try it out on Verathos Testnet:

https://t.co/lWdEJ586zp

Switch to Testnet using the selector in the top-right, find the Microtensor model, and put it to work.

More systems. More networks. More useful intelligence.

[Microtensor | Sn92 @Microtensor92]

Arena 5 is live: function calling

The fifth microtensor competition opens on subnet 92: **support / mt-3g**, round 1,240.

Opens in 9 hours

The task: A user request arrives with a list of available functions. The model emits the call or calls that answer it: correct function name, correct arguments, correct types. 200 tasks per round.

The problem it solves: Every AI agent is a loop of deciding which tool to use next, and that decision happens on every turn. At a million conversations a month, hosted API costs run $15,000 to $75,000 while a self hosted small model runs $150 to $800 for the same volume. The routing call is the highest frequency model call in any agent, and the one where frontier pricing makes least sense.

It is also where privacy bites hardest. The tool schemas describe internal systems and the arguments carry customer data, which is exactly what legal and compliance teams block from leaving the network.

Why a small model can win here: Tool use accuracy depends on schema adherence and argument correctness rather than parameter count. A purpose built 1B scores nearly three times a general 1B on the same benchmark. Tool calling is a trainable skill, not a byproduct of scale.

The envelope

disk 1.5 GB
memory 3 GB
p95 latency 8,000 ms

Base model: `Salesforce/xLAM-2-1b-fc-r`, purpose built for function calling and top of its class on the Berkeley leaderboard. Pinned at `6870fcf102f434bd01819302d85780b8e767925d`.

Metric: Tool call F1.

One difference from previous arenas: `chat = True` on this track, unlike code and classify which run raw completion.

The blocks:

opens 9,023,200
submissions close 9,030,400
evaluation ends 9,037,600

Opens 8th September 2:00pm UTC, closes September 9th 2:00pm UTC, evaluation ends Thursday 2:00pm UTC.

What changed since arena 4.

- Cost is the p95 total task latency for every system. Time to first token no longer sets the cost, so an empty answer cannot buy a cheap position.
- A profile that returns no output fails the system instead of recording zero cost.
- The reveal window after close is 300 blocks, about an hour, and validators read the chain once more after the deadline before rejecting anyone.

Submissions are sealed. Artifacts stay encrypted until the round closes, so nothing readable exists while the window is open. No reveal within the window means no score.

Arena details: https://t.co/QGr4pZlpWy

Sharpen your systems.

[Microtensor | Sn92 @Microtensor92]

AI is getting bigger. The hardware most people deploy on isn't.

The next phase won't be decided by which model tops a general leaderboard. It'll be decided by whether a model built for one task fits the memory, latency and cost budget a team already has.

That's what Microtensor is for. A network where independent researchers compete to build specialised models, and every one is measured rather than claimed.

→ Built for one task, not general ability
→ Measured on disk, memory, latency and CPU cores
→ Figures produced by independent validators on reference hardware
→ Signed certificates, so you verify rather than trust the builder
→ Download the weights and run them yourself
→ No API, no hosting, no account

Know a specialised model clears your bar, on your hardware, before you commit to it.

Explore → https://t.co/8Q0GMZnLkc

#Microtensor #SmallLanguageModels #EdgeAI #Bittensor #AIInfrastructure #TAO

[Microtensor | Sn92 @Microtensor92]

Round 1,239 has settled. Arena 4 ("classify / mt-3g") is complete.

62 systems entered. 42 were measured by all three validators, 16 never revealed and were dropped, and 4 recorded a zero cost and were unscored under the guard. The coordinator verified and signed the settlement at 07:26 UTC, and validators set weights at block 9,014,064.

Rewarded, by exclusive hypervolume

hotkey quality cost share
1 5CkuRmNC… 1.000 848 ms 30%
2 5FUaxZie… 0.956 8 ms 20%
3 5DoaV8JK… 0.989 496 ms 14%
4 5EqFqFeD… 0.850 3 ms 11%
5 5EnrDrZg… 0.950 177 ms 9%
6 5GnNdM1w… 0.939 225 ms 7%
7 5EyYXepK… 0.972 664 ms 5%
8 5CLfzJdv… 0.972 686 ms 4%

Read this round's cost with care. Costs here are time to first token, which is why single digit ms rows rank. Round 1,240 moves cost to total task latency and fails empty profiles (#48), and the reveal window rule will be stated up front.

The Registry is live:
https://t.co/8Q0GMZnLkc

Every certified system from rounds 1,237 to 1,239 is listed in one place. Each card shows what the validators measured, not what the miner claimed: quality, p95 latency, disk and peak memory on the reference hardware, with a seal for systems whose settlement was signed.

Open a card for the system view: its point on the cost and quality frontier, the signed certificate with the reports root and settlement signature, the measured envelope and archived files, and how to download the GGUF and run it with llama.cpp.

"Download weights" pulls the exact bytes the validators measured from https://t.co/OAU7igOSGd, so what you run is what was certified.

Leaderboard:

https://t.co/6VtlVAqUQV

[Microtensor | Sn92 @Microtensor92]

Round 1,239 submissions closes soon (block 9,006,601)

The classify/mt-3g window is still open. Commit now if you plan to enter. Anything that lands after block 9,006,601 is not admitted.

What happens at close:
• The roster freezes and sealed commitments are revealed.
• Evaluation runs on the frozen set until block 9,013,801.
• Results and rewards publish when the round settles after that.

Sealed submitters:

A sealed commitment is only a hash until you reveal it. **No reveal, no score.** Keep your miner running through the close block so the reveal goes out on time.

Arena: https://t.co/PTPckuamLW

[Microtensor | Sn92 @Microtensor92]

Thanks @csouthai. Specialised SLMs are what the network produces. Compression is just the constraint that forces the specialisation.

Revenue comes from Microtensor Compute, miner supplied GPUs doing four things. Serving the certified models through an inference API, sold per token or as reserved capacity. Model shaping and certification for people who bring their own model. Renting the hardware as clusters, sandboxes and storage. And idle capacity mines other subnets, with those earnings funding buybacks

[Microtensor | Sn92 @Microtensor92]

A 0.40 GB model, running on a single CPU core, scored 53.7% Pass@1 on BigCodeBench Instruct.

GPT-4o scored 51.1% on the same benchmark.

A model that fits in a fraction of a gigabyte outscored GPT-4o🧵

[Microtensor | Sn92 @Microtensor92]

Microtensor is expanding beyond certified inference.

We're introducing Microtensor Compute, a network of GPU capacity supplied by miners.

The idea: compute shouldn't sit idle. The same hardware pool moves between four workloads depending on where demand is.

1. Certified Inference serving the systems Microtensor already measures. API access, quality-floor routing, batch, reserved throughput, or dedicated hardware.

2. Subnet Mining capacity with nothing else to do mines other Bittensor subnets, instead of sitting idle.

3. GPU Rentals clusters, sandboxes, and managed storage, sold directly.

4. Model Shaping bring your own model for fine-tuning, training, or evaluation on certified reference hardware, with a signed certificate.

One pool, four demands. Rentals are honoured first, inference is served on arrival, model-shaping gets scheduled around both, and subnet mining absorbs whatever's left which is what keeps utilization high.

Microtensor Compute is coming, Stay tuned.

[Microtensor | Sn92 @Microtensor92]

The Microtensor model registry is live

https://t.co/8Q0GMZoj9K

Every system that has earned on the network is now browsable in one place, with the measurements that got it there attached.

Most model catalogues let you filter by parameter count and leave you to guess whether it will run. This one lets you filter by what you actually have: disk, memory, latency, CPU cores. Set your constraints, see what fits.

Nothing sits behind an API. No hosting, no account, no rate limit. Download the weights and run them on your own hardware.

Every entry carries a signed certificate, so what you get is the measured envelope, quality and cost, bound to the exact artifact that produced them. Not the builder's word for it.

Performance isn't promised. It's measured.

[Microtensor | Sn92 @Microtensor92]

Small models are usually benchmarked on general ability.

But deployment doesn't happen on an abstract benchmark.

It happens on your hardware, with your memory budget, your latency requirements and your cost constraints.

Microtensor measures systems under those conditions.

Performance isn't promised.

It's measured.

Leaderboard: https://t.co/1fmLG4I4m3

BigCodeBench: https://t.co/eazpCeDags

https://t.co/WMifOX8MvN $SN92

[Microtensor | Sn92 @Microtensor92]

None of these miners were simply told to maximize benchmark accuracy.

They were given a hardware envelope and competed to build the best system they could fit inside it.

The result is a different way of looking at model performance:

Not just how capable is the model?

But how much capability can you deliver within a real constraint?