BotPsy: How Tether AI Packed a Full Agentic Assistant Into 0.8 Billion Parameters
Tether AI Research's BotPsy is a 0.8B-parameter on-device agent that tops every sub-1.2B model on six tool-use benchmarks - and even beats a model 2.5× its size on five of them.
The race to shrink capable AI agents down to phone- and laptop-class hardware just got a serious entrant. On November 11, 2026, Tether AI Research released BotPsy, an open-weight, 0.8-billion-parameter language model purpose-built for on-device agentic workflows. The result challenges a prevailing assumption in the field: that useful tool-calling agents require billions of parameters and a cloud connection.
What BotPsy Is
BotPsy is a personal assistant that orchestrates software tools to carry out everyday tasks. Its target domains span web search and research, travel and lifestyle planning, commerce and finance, business and customer operations, device and vehicle control, file and document management, calendar scheduling, and email and messaging 1.
The model ships in multiple quantization formats - bf16, Q8, Q4, and GGUF - so it can run fully locally on consumer hardware without phoning home to any cloud API 3.
A Four-Stage Training Recipe
The engineering story behind BotPsy's performance is its post-training pipeline. Starting from Qwen3.5-0.8B, Tether applied four sequential stages 4:
- Supervised Fine-Tuning (SFT) on tool-use conversations to establish basic function-calling behavior.
- Multi-Teacher On-Policy Distillation (MOPD) using specialist Qwen3.5-4B teacher models, transferring structured agentic knowledge into the smaller student.
- Two rounds of Reinforcement Learning (RL) that optimize tool calling, instruction following, general reasoning, and structured output quality.
This cascade - SFT → distillation → RL × 2 - is what allows such a compact model to handle multi-step, multi-tool workflows that would ordinarily demand a much larger parameter budget.
Benchmark Results: Beating Models 2.5× Its Size
The headline numbers are striking. BotPsy outperforms every tested model in the ≤1.2B-parameter weight class - including Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think, and MiniCPM5-1B-Think - across all six tool-use and agentic benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus 1.
More impressively, it beats Qwen3.5-2B-Think (a model 2.5× its size) on five of those six benchmarks, posting scores of 44.0 on BFCL v4, 58.7 on ACEBench, 51.5 on τ-bench, 64.6 on ToolSandbox, and 4.5 on BrowseComp-Plus 2. The one exception is Claw-Eval, where the 2B model edges ahead 48.9 to 46.5 6.
The gains relative to the base model are especially pronounced on multi-turn tasks. On BFCL v4 multi-turn, BotPsy scores 46.1 versus 7.5 for the base model - and significantly outpaces the 2B model's 23.8 8. General reasoning also improved: GPQA Diamond rose from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9 8.
Instruction-following saw the largest delta of all: IFEval jumped from 59.7 (base) to 90.8, surpassing the 2B model's 80.0 4.
Efficiency Gains on Device
Performance isn't the only win. BotPsy uses approximately three times fewer reasoning tokens than its base model on GPQA Diamond, and roughly 2.7× fewer on MMLU-Pro, with fewer repetitive reasoning loops and fewer runs hitting long-context limits 2. For on-device deployment, where memory and battery matter, that token efficiency is as important as raw accuracy.
Limitations and Honest Trade-offs
Tether is candid about where BotPsy falls short. The model works best on single-call and multi-step tasks in communication, commerce, services, and device management. It remains weaker on open-ended research tasks, some web-search and memory-intensive workflows, lower-resource languages (Hindi, Russian), and complex banking workflows 2.
No independent evaluation or real-world deployment data has been published yet, so the benchmark scores, while encouraging, haven't been stress-tested outside Tether's own harness 9.
The Bigger Picture: Private, Local AI Ecosystems
Tether frames BotPsy as more than a one-off model release. The company describes it as a step toward a private, local AI ecosystem, with future work aimed at encrypted peer-to-peer search, stronger local fine-tuning, secure agent-to-agent communication, and offline-to-online task delegation 4. The underlying thesis - that privacy-preserving agents running on personal hardware represent an alternative to expensive, opaque cloud-dependent models - is a coherent product vision, not just a research curiosity.
Why It Matters
Sub-billion-parameter models capable of real agentic workflows change the deployment calculus for personal AI significantly. If a model this small can reliably handle calendars, emails, file management, and multi-turn web research without a data center, the barriers to private, low-latency, low-cost AI assistants drop considerably.
BotPsy won't replace frontier reasoning models for complex analytical tasks, but that's not its job. As a specialist for routine device-side automation, it demonstrates that targeted post-training - the right combination of distillation and reinforcement learning - can unlock capabilities from small models that most practitioners assumed required an order of magnitude more parameters. For engineers building edge AI stacks, that's a result worth watching carefully.
Sources
- 1. BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model (Hugging Face)
- 2. Tether AI Research Releases BotPsy 0.8B On-Device Agentic Model (Hyper.ai)
- 3. BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic (TechBytes)
- 4. Could A Small Language Model Handle Your Personal Assistant Tasks? (Toksick Magazine)
- 5. Tether introduces an 800-million-parameter AI assistant built for on-device tasks (Superpower Daily)
This article was researched and drafted by an AI writer agent (claude-sonnet-4-6) and reviewed by an editor agent before publishing.