Full-stack AI engine built on Bittensor. NexisGen SN70 | Leoma SN99 | Eirel SN36
Bittensor
Joined March 2026
- Tweets11
- Following5
- Followers172
- Likes7
Rendix AI retweeted
Who's already using Eirel?
We want feedback from real users.
How is the speed? How is the research quality? Are the models routing well? Is the agent workflow useful enough for daily work?
Another huge update is coming.
We’re improving latency, tool execution, model selection, research depth, and overall reliability.
In a few months, Eirel won't just be an alternative.
It will be the default.
Rendix AI retweeted
Eirel has reached 500 users just a few days after launch.
This is only the beginning.
More users are discovering what we are building on Bittensor SN36: deep research, image generation, web search, and multi-model AI in one product.
I believe more users will come, use it, and pay for it.
Real product. Real users. Real growth.
Thanks to $TAO miners.
Rendix AI retweeted
We are on building a multi-turn conversational assistant across instant and thinking modes.
It is backed by owner-routed tool services, including web search, URL fetching, Python sandbox execution for verifiable computation, and RAG retrieval over per-run document corpora.
We are now building the product and will soon release it with the agent code from the top miner.
Stay tuned for the updates.
#Bittensor $TAO
Rendix AI retweeted
Since yesterday, miners have added 5 more hours of video generation data to NexisGen.
That brings the total to 27 hours of video generation data now available.
You can check our progress here:
nexisgen.ai/catalogue/video_…
Thanks to all the miners contributing.
#Bittensor $TAO
Rendix AI retweeted
We are starting with AI training datasets for video generation models, but this is only the first step.
The current dataset includes structured video clips like this:
- 1280×704 resolution
- 24 FPS
- 121 frames per clip
- structured clip IDs, producers, and intervals
- link: nexisgen.ai/catalogue/video_…
In just a few days, we’ve already collected around 22 hours of video data, with a lot more coming soon.
SN70 is building the data layer for decentralized AI training, starting with video.
#Bittensor $TAO
Rendix AI retweeted
All three @rendix_network subnets are live now.
Subnet 70 - @nexisgen_ai is to get the large scale AI Training Data for various models. (xitter.gq/nexisgen_ai/status/205…)
Subnet 99 - @leoma_ai to build the best Video Generation Model. (xitter.gq/leoma_ai/status/205278…)
Subnet 36 - @eirel_ai is the execution layer of multimodal AI workflows. (xitter.gq/eirel_ai/status/205494…)
All three subnets are live and moving fast in parallel.
The product needs high-quality output, and that comes from both a strong agent and a strong model.
A strong model comes from a good training pipeline and high-quality data.
That’s why we need all three subnets working together to build a product that can compete with SOTA.
#Bittensor $TAO
Rendix AI retweeted
Incentive Mechanism updates!
Reward miners whose video datasets produce the best fine-tuned TI2V models. Quality is measured by training a LoRA on each miner's data and scoring its generations with VBench.
How a cycle works:
- Miners upload 400-clip datasets (1280×704, 24fps, 121 frames, captioned) to their own R2 buckets.
- Owner-trainer picks eligible miners, validates the dataset spec, trains a LoRA on each, and publishes eval-prompt videos to a shared bucket.
- Validators score every miner's videos with VBench across 8 dimensions, sign and submit scores to the subnet API.
- API averages scores across validators → total_score.json.
- Set-weight (every 300 blocks): top-5 miners get weights [1, 1/2, 1/4, 1/8, 1/16] normalized; no scores → burn to UID 0.
VBench dimensions: i2v_subject, i2v_background, subject_consistency, background_consistency, motion_smoothness, dynamic_degree, aesthetic_quality, imaging_quality Aggregate = mean of the 8.
Dataset must-haves (any failure → rejected):
- Exactly 400 clips, 1280*704, 24fps, 121 frames
- Non-empty per-clip caption
- No internal (source_url, start_sec) duplicates
- <=100 clips overlap with the global overlap index
Anti-gaming:
- Selected miners auto-marked invalid; re-entry requires a top-5 finish.
- Global overlap index blocks resubmitting the same URLs.
- Multiple independent validators average each score.
Links:
- Dashboard: nexisgen.ai/floor
- Github: github.com/RendixNetwork/nex…
bittensor:native Bittensor