Bottensor is a research lab. Five open-weight models live on HuggingFace today. Current focus: reinforcement learning for multi-step agents in verifiable environments.
Every model in the family starts as a non-player character — a background role in a world where frontier labs train the protagonists. The work is to level them up. Proprietary data, logic-tree reasoning, domain-specialist training. The bet is that small models, trained right, beat generalist protagonists inside their vertical. NPC is the family. Bottensor is the lab building them.
What we're working on next.
Current focus: reinforcement learning for agents. Post-training language models to plan and act over multiple steps in verifiable environments — where every reward is a deterministic check against world state, not a judge model.
Agent World is the testbed: a deterministic multi-agent environment with geometric rewards, an LLM tool-call interface, and full replay. Current findings from the prompted baselines: a 0.5B model speaks the action protocol but can't plan (48% invalid actions); a 7B model is near protocol-perfect (0.9% invalid) yet completes zero deliveries itself. Planning, not protocol, is the bottleneck — the RL training phase targets exactly that gap. Explore Agent World →
Watch agents work. Then try it yourself.
Agent World is the deterministic multi-agent environment behind our preregistered RL work: a 2D office where agents fetch crates and deliver them to a goal zone, with exact geometric rewards and byte-identical replays. Watch the scripted baseline solve it, watch a tiny LLM wander, watch a 7B politely do nothing while its teammate works — then play a round against the scripted agents in your browser.
Five models shipped.
Three in research.
Earlier work on small-model training and verifiable reasoning — including honest null results — that led to the current agents focus. All open, all on HuggingFace: weights, recipes, and evaluations are public as they land.
32B finance reasoning specialist. Fine-tuned on curated market examples. Weighs evidence, flags risks, delivers structured theses. Base: Qwen2.5-32B-Instruct. Method: QLoRA SFT → merged → GPTQ 4-bit. CryptoQA 93.6%.
7B process reward model. Verifies NPC Fin’s reasoning step by step. The alignment layer that keeps the specialist honest. Spearman 0.9234 · F1 0.8421.
1.7B lightweight model for routing, lookups, translation, and general-purpose tasks. Fast inference, compact footprint. Shipped in both safetensors and GGUF formats.
7B agentic specialist for tool use, planning, and multi-step execution. Base: Qwen2.5-7B-Instruct. Method: QLoRA SFT → merged → GPTQ 4-bit + GGUF. Shipped in safetensors, LoRA, GPTQ, and GGUF formats.
The first from-scratch model in the NPC family. 501M-parameter decoder-only transformer (24 layers, 1024 hidden, GQA 16/4, SwiGLU, RoPE, RMSNorm) trained from random initialization on 8.93B tokens of curated open-source data on a single A40. Released as base, SFT, and GGUF variants alongside a methodology paper documenting the post-training capability ceiling at sub-1B parameters.
Production-grade code generation across Python, TypeScript, Solidity, Java. Repo-level context and tool-use.
Multi-step reasoning trained with GRPO against a SymPy-verified reward — the verifiable-rewards line that led to the current agents focus.
Extended context for document and codebase reasoning. Target context length TBD.
NPC Fin
A 32B finance specialist fine-tuned on curated market data, quantized for fast inference, and paired with a 7B process reward model that verifies reasoning step by step. Open weights, open recipes.
Preprints & recipes.
Open preprints on Zenodo. Each paper documents the recipe behind a shipped model — data, training, evaluation. Cited as you would any preprint.
Packages & tooling.
The research stack, published. npm and PyPI libraries that grew out of the NPC work — installable, open, and used in production.
A small research lab.
Bottensor is a small research lab. We build small, fast, specialized AI models for problems generalists can't solve well. The NPC Model Family is our long-term project — one model per real-world domain, shipped with open weights and open recipes.
We run the whole pipeline end-to-end: data curation, fine-tuning with QLoRA and Unsloth, quantization, and evaluation. Roughly 25% of what we do stays closed (proprietary datasets, training recipes we're still refining), and 75% ships open (weights, code, evals). That ratio will shift as the research matures.
Founded by Rama Krishna Bachu.
Building the NPC Model Family end-to-end. Data, training, evaluation, research direction. Previous: 7+ years software engineering, MS Computer Science.
Built with
End-to-end AI infrastructure — from data pipelines to production inference.