Product
The SLM Ensemble
One focused product. Eight C++ specialists. No marketing fluff, just the models we actually train, evaluate, and ship.
Language Models
SLM Ensemble
Specialists beat generalists. Always.
The scaling hypothesis is dead—long live specialists. An ensemble of 8 models (4B-8B params each, 0.8B-1.6B active) that outperforms 70B generalists on C++ at a fraction of the cost. Mamba 3 + Transformers hybrid architecture, trained with Muon on 100-200B tokens per specialist, served in NVFP4 on Blackwell and GB10.
8
Specialists
4B-8B
Params (0.8-1.6B active)
100-200B
Tokens each
What you get
- 8 specialists: algorithms, templates, memory, concurrency, systems, build, debug, STL
- Mamba 3 + Transformers hybrid layer stack
- Muon optimizer, FP16/BF16 training
- NVFP4 inference on Blackwell (B200) and GB10
- 100-200B tokens per specialist, C++ curriculum
- gdb / rr ground-truth integration, not vibes
Want the technical details?
Architecture, training recipe, data pipeline, evaluation, NVFP4 inference, and honest GB10 war stories — all grounded in our working stack.