Invergent

We cater to a wide variety of industries and use cases, from digital transformation to big data, custom LLMs, NLP applications, AI agents, churn prediction, tim

[Invergent @invergentai]

No component assumes anything about the others beyond a small set of interfaces. This means if the harness crashes, a new one can pick up where it left off by replaying the session log. If the sandbox dies, the harness can provision a new one and continue. The API server is the only component with access to tenant-level credentials (DB, S3) and issues scoped tokens to the workers, which in turn manage sandbox credentials.

[Flavius Burca @flaviusburca]

Introducing the open platform for running Managed Agents at scale, bringing Claude Managed Agents on-premise.

- Built around a clear separation between reasoning (“brain”) and execution (“hands”).
- Multi-tenant, Multi-user
- Enterprise-grade security
- Scales massively to thousands of agents / sessions / users

https://t.co/yIfs2ykYWe

[surogate @surogate_ai]

We just shipped FlashOptim-style 8-bit AdamW in Surogate 🔥

Replaced bitsandbytes' quantile-map approach with softsign/sqrt transforms for optimizer state quantization.

Results on 50M params:
• 27x lower quantization error (0.012% vs 0.33% relative error)
• 1.24x faster kernel execution
• Same memory footprint (~2 bytes/param)

Better optimizer state fidelity → less training drift → better final model quality, especially on long runs.

Drop-in replacement — no config changes needed.

[surogate @surogate_ai]

https://t.co/Io5KaNoFVF

[Invergent @invergentai]

Kudos @DAlistarh for their NVFP4 improvement in Quartet-II.

https://t.co/8xQxfAhQhu

The nvfp4-quartet recipe is now available in Surogate.

45k TPS on RTX 5090 with Qwen3-0.6B vs 35k TPS with BF16

[Invergent @invergentai]

Surogate now supports NorMuon+AdamW for pre-training and fine-tuning:

- 1D params / embeddings / norms use AdamW
- 2D weight matrices use NorMuon

Here is how NorMuon algorithm for 2D weights is implemented:

- Momentum update with 8-bit quantization
- Polar Express orthogonalization (Newton-Schulz iterations)
- Variance reduction (Adafactor-style)
- Cautious weight decay (sign-aware)
-LR scaling by sqrt(max(1, M/N)) based on weight shape

Results: faster training convergence for pre-training and FFT

[Invergent @invergentai]

Google uses C++
Meta uses C++
Microsoft uses C++
Amazon uses C++
Apple uses C++
Netflix uses C++
Adobe uses C++
NVIDIA uses C++
Intel uses C++
Uber uses C++
Dropbox uses C++
Cloudflare uses C++
Spotify uses C++
...

[Invergent @invergentai]

This is what successful MoE fine-tuning is supposed to look like.

Controlling 'router_aux_loss_coef' and 'router_z_loss_coef' is essential to prevent expert load imbalance and router collapse.

Surogate exposes unique MoE controls and metrics end-to-end. Because if you can’t see router health and expert load in real time, you’re already too late.

https://t.co/odCy0cHpST

[Invergent @invergentai]

In the GenAI race, iteration speed is the only metric that truly matters.

If your training loop takes 3 days, you iterate twice a week. If it takes 3 hours, you iterate twice a day. That compound effect is massive.

That’s why our tagline is simple: 

Insanely fast LLM training.

Surogate is engineered to minimize overhead and maximize velocity. It’s not just about raw compute, it’s about getting from data to model faster than anyone else.

⚡️ Why switch to Surogate?

•Optimized data loading.

•Reduced training overhead.

•Built for fine-tuning at scale.

•Start training at the speed of thought.

Join the high-performance era:
👉 GitHub: https://t.co/m64EyDDq49
👉 Website: https://t.co/fsydVcjOf8
👉 X: @surogate_ai

#GenAI #Speed #FineTuning #DevTools #InvergentAI #LLM #LLMTraining

[Invergent @invergentai]

AI compute is likely one of your fastest-growing cost centers. But the real cost isn't just the hardware - it's the inefficiency.

Every hour your models spend training unnecessarily is an hour of delayed product launch and wasted OpEx.

Introducing Surogate Enterprise.

We built the enterprise tier for organizations that treat AI as a core business driver, not an experiment. We help you extract maximum value from your compute infrastructure, drastically improving your AI ROI.

Why Surogate Enterprise?

📈 Cut OpEx: Maximize hardware utilization to do more with less.

🚀 Accelerate Time-to-Market: Faster iteration cycles mean faster product releases.

🛡 Enterprise Readiness: Scalability, support, and reliability you can build a business on.

Stop treating training as a bottleneck. Make it your competitive advantage.
Let’s discuss your infrastructure strategy.

Join the high-performance era:
👉 GitHub: https://t.co/m64EyDDq49
👉 Website: https://t.co/fsydVcjOf8
👉 X: @surogate_ai

#AIStrategy #ROI #EnterpriseAI #Invergent #Leadership #TechBusiness #LLM

[Invergent @invergentai]

Most training libraries leave your GPU idling. We fixed that.

The bottleneck in LLM training often isn't the hardware itself - it's the software abstraction layers that prevent you from using it fully.

At Invergent, we built Surogate to strip away the inefficiencies. It’s designed to push your training and fine-tuning processes to the absolute practical limits of your infrastructure.

Whether you are running on a single GPU or a cluster, Surogate ensures you are getting every FLOP you paid for.

Don't let your hardware sleep.

Join the high-performance era:
👉 GitHub: https://t.co/m64EyDCSeB
👉 Website: https://t.co/fsydVcjgpA
👉 X: @surogate_ai

#LLM #Invergent #OpenSource #LLMTraining #FineTuning #MachineLearning #GPUOptimization

[Kimi.ai @Kimi_Moonshot]

🚀 Hello, Kimi K2 Thinking!
The Open-Source Thinking Agent Model is here.

🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%)
🔹 Executes up to 200 – 300 sequential tool calls without human interference
🔹 Excels in reasoning, agentic search, and coding
🔹 256K context window

Built as a thinking agent, K2 Thinking marks our latest efforts in test-time scaling — scaling both thinking tokens and tool-calling turns.

K2 Thinking is now live on https://t.co/YutVbwktG0 in chat mode, with full agentic mode coming soon. It is also accessible via API.

🔌 API is live: https://t.co/EOZkbOwCN4
🔗 Tech blog: https://t.co/n7xxaszqzF
🔗 Weights & code: https://t.co/4ukcXB0iP6

[Invergent @invergentai]

Cloud AI has a secret: you don't control your data, and you can't predict your costs
We're changing the game
Introducing DenseMAX, the complete AI ecosystem that runs in your infrastructure, powered by DenseMAX Studio, a no-code ai platform
#AI #DataSovereignty #LLMOps