BOTTENSOR
Bottensor — Open Research Lab
Non-player characters,
leveling up.

Bottensor is a research lab. Five open-weight models live on HuggingFace today. Current focus: reinforcement learning for multi-step agents in verifiable environments.

View on HuggingFaceRead the research
5
Shipped models
3.2K
HF downloads · 30d
32B
Largest model
Apache 2.0
License
Why NPC

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.

Open Research LabNPC Model FamilySmall SpecialistsQwen2.5-32BSmolLM2-1.7BUnslothQLoRAGPTQ 4-bitProcess Reward ModelsHuggingFaceOpen WeightsApache 2.0Fine-TunedOpen Research LabNPC Model FamilySmall SpecialistsQwen2.5-32BSmolLM2-1.7BUnslothQLoRAGPTQ 4-bitProcess Reward ModelsHuggingFaceOpen WeightsApache 2.0Fine-Tuned
Research

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 →

Research collaboration
Get in touch to discuss the thesis or the roadmap.
research@bottensor.xyz →
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.

Watch the replaysPlay it
7 JSON tool-call actions · verifiable geometric rewards · deterministic replay · in-browser demo (the Python env is canonical)
Total HF downloads
across 15 repos · all time
3.2K
NPC Fast
2 repos
952
29.9%
NPC Agent
4 repos
985
30.9%
NPC Nano
4 repos
514
16.1%
NPC Fin
1 repo
51
1.6%
NPC Fin PRM
1 repo
62
1.9%
Published Work

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.

01
Shipped
NPC Fin
Finance · 32B

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%.

51 downloads · 32B · QLoRA SFT · updated Apr 2026
HuggingFace →
02
Shipped
NPC Fin PRM
Alignment · 7B

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.

62 downloads · 7B · Process RM · updated Apr 2026
HuggingFace →
03
Shipped
NPC Fast
Lightweight · 1.7B

1.7B lightweight model for routing, lookups, translation, and general-purpose tasks. Fast inference, compact footprint. Shipped in both safetensors and GGUF formats.

952 downloads · 1.7B · bf16 + GGUF · updated Apr 2026
HuggingFace →
04
Shipped
NPC Agent
Agentic · 7B

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.

985 downloads · 7B · QLoRA SFT · updated May 2026
HuggingFace →
05
Shipped
NPC Nano
From-scratch · 0.5B

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.

493 downloads · 0.5B · From-scratch · updated May 2026
HuggingFace →
06
Research
NPC Coder
Code Generation

Production-grade code generation across Python, TypeScript, Solidity, Java. Repo-level context and tool-use.

07
Research
NPC Reason
GRPO · SymPy Verifier

Multi-step reasoning trained with GRPO against a SymPy-verified reward — the verifiable-rewards line that led to the current agents focus.

08
Research
NPC Context
Long Context

Extended context for document and codebase reasoning. Target context length TBD.

Flagship Research Artifact

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.

51 HF downloads (30d)· updated Apr 2026View on HuggingFace →
model_spec.json
ModelNPC Fin 32B
BaseQwen2.5-32B-Instruct
MethodQLoRA SFT → Merged → GPTQ 4-bit
Parameters32 billion
QuantizationGPTQ 4-bit (≈19 GB)
Training Data~32K examples · ~60M tokens
Context Length32,768 tokens
LicenseApache 2.0
Training Data Distribution
market_signal35%
finance_general25%
logic_tree20%
macro12%
cross_market8%
Open
NPC Fin 32B SFT
Fine-tune · QLoRA
1 downloads · npc-fin-32b-sft
Open
NPC Fin PRM
Process reward model · 7B
4 downloads · npc-fin-prm-7b
Open
NPC Fast
Lightweight · 1.7B
15 downloads · npc-fast-1.7b
Open
NPC Fast GGUF
GGUF quants · llama.cpp
10 downloads · npc-fast-1.7b-gguf
Open
NPC Agent
Fine-tune · QLoRA
15 downloads · npc-agentic-7b-v3
Open
NPC Agent LoRA
LoRA adapter
4 downloads · npc-agentic-7b-v3-lora
Open
NPC Agent GPTQ
GPTQ 4-bit
6 downloads · npc-agentic-7b-v3-gptq-4bit
Open
NPC Agent GGUF
GGUF quants · llama.cpp
41 downloads · npc-agentic-7b-v3-gguf
Open
NPC Nano Base
From-scratch · 0.5B
12 downloads · npc-nano-0.5b-base
Open
NPC Nano SFT
SFT · 0.5B
19 downloads · npc-nano-0.5b-sft
Open
NPC Nano GGUF
GGUF quants · llama.cpp
22 downloads · npc-nano-0.5b-sft-gguf
Publications

Preprints & recipes.

Open preprints on Zenodo. Each paper documents the recipe behind a shipped model — data, training, evaluation. Cited as you would any preprint.

Preprint · CC-BY-4.0
NPC Fast 1.7B: Building a Usable Small Model on a Single H100
doi.org/10.5281/zenodo.19771040Zenodo · open access
Preprint · CC-BY-4.0
Cheap PRMs: Training a Process Reward Model on a Single H100
doi.org/10.5281/zenodo.19800784Zenodo · open access
Preprint · CC-BY-4.0
NPC Fin 32B: Multi-GPU QLoRA on 12×H100
doi.org/10.5281/zenodo.19802598Zenodo · open access
Preprint · CC-BY-4.0
NPC Agentic 7B v3: Single-GPU QLoRA Recipe
doi.org/10.5281/zenodo.19954103Zenodo · open access
Preprint · CC-BY-4.0
NPC Nano 0.5B: From-Scratch Pretraining and the Post-Training Capability Ceiling at Sub-1B Parameters
doi.org/10.5281/zenodo.20349362Zenodo · open access
ORCID 0009-0000-1298-0681· Bottensor (Independent Research)· All papers CC-BY-4.0
Open Source

Packages & tooling.

The research stack, published. npm and PyPI libraries that grew out of the NPC work — installable, open, and used in production.

Package downloads
npm
3 packages · last year
694
@bottensor/cogito19728%
@bottensor/engram15823%
@bottensor/forge33949%
PyPI
3 packages · last month · as of Jun 2026
1,188
reasoning-bank26422%
bottensor-fleet60551%
polyrt31927%
Windows differ by registry — npm reports last-year, PyPI reports last-month. Shown separately, not summed.
npm197
@bottensor/cogito
197 downloads · last yearnpm
npm158
@bottensor/engram
158 downloads · last yearnpm
npm339
@bottensor/forge
339 downloads · last yearnpm
PyPI264
reasoning-bank
264 downloads · last month · as of Jun 2026PyPI
PyPI605
bottensor-fleet
605 downloads · last month · as of Jun 2026PyPI
PyPI319
polyrt
319 downloads · last month · as of Jun 2026PyPI
npm counts are last-year totals · PyPI counts are last-month (pypistats) · refreshed weekly
About

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.

R
Rama Krishna Bachu
Founder · Bottensor

Building the NPC Model Family end-to-end. Data, training, evaluation, research direction. Previous: 7+ years software engineering, MS Computer Science.

Stack

Built with

End-to-end AI infrastructure — from data pipelines to production inference.

Qwen2.5-32B
Unsloth
QLoRA
vLLM
Python
PyTorch
GPTQ 4-bit
HuggingFace
OpenAI API
Next.js
TypeScript
MongoDB
RunPod
A100 / H100
Vercel