Tether AI Releases BotPsy 0.8B Agent Beating 2B Models on Benchmarks

Tether AI releases BotPsy 0.8B model

Tether AI Research just dropped BotPsy, a 0.8-billion-parameter model built specifically for on-device agentic workflows and function calling. It landed on November 11, 2026, branching off Qwen3.5-0.8B. The target list is practical: web search, travel planning, commerce, device control, calendar management, and messaging. Tether pitches it as a personal assistant that keeps your data local, dodging the cost, privacy headaches, and vendor lock-in of frontier cloud APIs. Weights are out under Apache 2.0 for research and education, served up in bf16, Q8, and Q4 precisions as Hugging Face safetensors and GGUF files ready for the QVAC SDK and llama.cpp.

BotPsy benchmark results vs larger models

The benchmarks are where things get interesting. BotPsy roughly doubles the agentic chops of its base across three core suites: BFCL v4 jumps from 23.4 to 44.0, ACEBench climbs 30.1 to 58.7, and τ-bench goes 23.2 to 51.5. More striking, it beats every model up to 1.2B parameters Tether tested-including Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think, and MiniCPM5-1B-Think-on all six tool-use and agentic benchmarks. It even tops Qwen3.5-2B-Think, a model two and a half times larger, on five of six: BFCL v4 (44.0 vs 37.7), ACEBench (58.7 vs 55.4), τ-bench (51.5 vs 46.5), ToolSandbox (64.6 vs 41.2), and BrowseComp-Plus (4.5 vs 2.8). Only on Claw-Eval does the 2B model nose ahead, 48.9 to 46.5.

Multi-turn reliability shows the sharpest gains. On τ²-bench, which stresses long customer conversations with policy adherence, BotPsy hits 82.5 in telecom against 69.3 for the 2B Qwen. On BFCL v4's multi-turn subset it leaps from 7.5 to 46.1, nearly double the larger model's 23.8. Grouped by application, it leads all tested models in commerce and services at 58.3 versus 47.4, and in device and system management at 46.1 versus 31.9. The model also shows strong restraint: on BFCL v4's hallucination tests-checking that a model declines when no appropriate tool exists-it scores 86.1, the highest of any model evaluated. On ACEBench's special cases (incomplete requests, wrong parameters, unsupported tools) it reaches 51.3, up from 13.7 for the base.

Instruction following improves dramatically. IFEval climbs from 59.7 to 90.8, and IFBench from 22.0 to 55.9, the highest IFBench score among all models tested. General knowledge doesn't suffer: GPQA Diamond rises from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9, with BotPsy topping every other model at or below 1.2B parameters on all four benchmarks. Tether notes knowledge still scales primarily with parameter count, so the model remains best suited to low- and medium-difficulty tasks and isn't intended for high-stakes use.

Safety metrics move in the right direction. In automated red-teaming across six harm categories, BotPsy produces unsafe responses 21 percent of the time, compared with 36 percent for its base and 35 to 47 percent for the other sub-1.2B models tested. Under prompt-injection attacks-a critical vector for agents that ingest untrusted tool outputs-the unsafe rate drops from 46 percent to 31 percent.

Efficiency gains are substantial. On questions all five compared models answer correctly, BotPsy uses about three times fewer reasoning tokens than its base on GPQA Diamond (3,690 versus 11,563) and 2.7 times fewer on MMLU-Pro (1,912 versus 5,105). Repetitive reasoning loops fall from 22.9 percent to 1.4 percent across IFBench attempts, and runs hitting the 32k-token limit drop from 28.3 percent to 0.2 percent. Shorter, more decisive reasoning should translate directly into lower on-device latency.

BotPsy training pipeline and feedback stages

The training pipeline spans four stages and incorporates three types of feedback, though Tether hasn't disclosed the full data composition. The company, formally Tether Data, S.A. de C.V., frames BotPsy as a foundation for expanding on-device agentic capabilities. Future work includes integrating Hypersearch for decentralized peer-to-peer queries that bypass corporate trackers, and optimizing QVAC Fabric for local fine-tuning that adapts to individual user preferences and specialized tool sets.

Running agentic workloads locally with BotPsy

For developers, the immediate takeaway is that a sub-billion-parameter model can now handle real agentic workloads, multi-step tool use, policy-constrained conversations, and reliable instruction following, while running locally on consumer hardware. The Apache 2.0 release means teams can inspect, modify, and deploy the weights without vendor lock-in. The GGUF quantizations make it practical to run on phones, laptops, or embedded devices via llama.cpp or the QVAC SDK. The trade-off is clear: BotPsy will still make mistakes, especially on multi-step information retrieval, and it isn't suited for high-stakes domains. But for the large class of routine, privacy-sensitive tasks that don't require frontier intelligence, it offers a compelling alternative to cloud APIs that can change pricing, deprecate endpoints, or expose user data.

Frequently asked questions

What is BotPsy and when was it released?

BotPsy is a 0.8-billion-parameter AI model released by Tether AI Research on November 11, 2026, built for on-device agentic workflows and function calling. It branches from Qwen3.5-0.8B and targets tasks like web search, travel planning, commerce, device control, calendar management, and messaging while keeping data local.

How does BotPsy perform compared to larger models on agentic benchmarks?

BotPsy beats every tested model up to 1.2B parameters on all six tool-use benchmarks. It also outperforms Qwen3.5-2B-Think (2.5x larger) on five of six benchmarks: BFCL v4 (44.0 vs 37.7), ACEBench (58.7 vs 55.4), τ-bench (51.5 vs 46.5), ToolSandbox (64.6 vs 41.2), and BrowseComp-Plus (4.5 vs 2.8). Only Claw-Eval favors the 2B model (48.9 vs 46.5).

What licensing and deployment options does BotPsy offer?

BotPsy weights are released under Apache 2.0 for research and education. Available formats include bf16, Q8, and Q4 precisions as Hugging Face safetensors and GGUF files, compatible with QVAC SDK and llama.cpp for on-device deployment on phones, laptops, or embedded devices.

What are BotPsy's key limitations?

BotPsy is best suited for low- and medium-difficulty tasks, not high-stakes use. Knowledge still scales primarily with parameter count. It can make mistakes on multi-step information retrieval and shows lower performance on complex reasoning compared to larger frontier models.

How does BotPsy improve safety and efficiency over its base model?

BotPsy reduces unsafe responses in automated red-teaming from 36% to 21% across six harm categories, and prompt-injection unsafe rates drop from 46% to 31%. It uses ~3x fewer reasoning tokens on GPQA Diamond (3,690 vs 11,563) and 2.7x fewer on MMLU-Pro. Repetitive reasoning loops fall from 22.9% to 1.4%, and 32k-token limit hits drop from 28.3% to 0.2%.

Source: Hugging Face