The Sub-Billion Agent Race Just Got Real
Tether AI Research dropped BotPsy, an 800M-parameter model built for function calling and agentic workflows. The pitch is simple: not every task needs frontier-scale intelligence, and you shouldn't have to ship your data to the cloud for routine agent work. BotPsy is a "personal assistant" that handles web search, travel, commerce, calendar, email, and device control—all on-device.
What makes this notable isn't just the sub-billion parameter count. It's that BotPsy beats every model up to 1.2B parameters on tool-use benchmarks, and on five of six benchmarks it outperforms Qwen3.5-2B-Think—a model 2.5× its size. The gap isn't close: on BFCL v4 it scores 44.0 versus 37.7 for the 2B sibling, on ACEBench 58.7 versus 55.4, and on τ-bench 51.5 versus 46.5.
This is the new battleground. While frontier labs push reasoning tokens and chain-of-thought at scale, a parallel race is happening for the smallest viable agent. BotPsy suggests that with the right training recipe, 800M parameters is enough to be genuinely useful.
What BotPsy Actually Does
BotPsy is fine-tuned on top of Qwen3.5-0.8B and roughly doubles its base model's agentic performance. BFCL v4 jumps from 23.4 to 44.0, ACEBench from 30.1 to 58.7, τ-bench from 23.2 to 51.5. The model specializes in domains like commerce and services (scoring 58.3 on multi-step tasks), device and system management (46.1), and sustained multi-turn customer service conversations.
On τ²-bench—where the agent must hold a long conversation, follow strict policy, and act on behalf of a customer—BotPsy scores 82.5 in telecom versus 69.3 for Qwen3.5-2B-Think. For multi-turn BFCL v4 tasks, it rises from 7.5 to 46.1 over base, nearly double the 2B model's 23.8.
The model also knows when not to act. On BFCL v4's hallucination tests (does the model decline when no appropriate tool exists?), it scores 86.1—the highest of all models tested. On ACEBench's special cases (incomplete requests, wrong parameters, unsupported tools), it scores 51.3 versus 13.7 for the base model.
Instruction following improves dramatically: IFEval rises from 59.7 to 90.8, IFBench from 22.0 to 55.9—the highest IFBench score in the weight class.
The Training Recipe: Four Stages, Three Feedback Types
Tether's approach uses a four-stage pipeline: supervised fine-tuning (SFT), multi-teacher on-policy distillation (MOPD), then two rounds of reinforcement learning.
Stage 1: Learning from Demonstrations
Full-parameter SFT on tool-use conversations—multi-turn agent trajectories from simulated environments, single/parallel/multi-step call examples, and negative cases where no tool should fire. This stage boosts tool use but degrades knowledge and instruction following below the base model.
Stage 2: Learning from Specialist Teachers
On-policy distillation: the student generates responses, and a specialist teacher provides per-token feedback via reverse KL divergence. Tether built three RL-trained Qwen3.5-4B specialists—one for tool use, one for general reasoning, one for instruction following—and routes each prompt to the appropriate teacher. This recovers the knowledge and instruction-following capabilities that SFT degraded, while also improving tool-use metrics.
Stage 3-4: Learning from Rules and Verifiers
Two rounds of reinforcement learning using DAPO, learning only from rewards computed by rules and verifiers. The first round focuses on tool calling with instruction-following rewards; the second adds general knowledge. By the end, mean scores across all three capability groups sit above the base model.
The result is a model that preserves general skills while specializing. GPQA Diamond rises from 33.2 to 38.1, MMLU-Pro from 41.6 to 48.9—outperforming every model ≤1.2B parameters that Tether tested.
Efficiency Gains: Shorter Reasoning, Fewer Loops
On questions all five comparison models answer correctly, BotPsy reaches the answer with roughly 3× fewer reasoning tokens than its base on GPQA Diamond (3,690 versus 11,563) and 2.7× fewer on MMLU-Pro (1,912 versus 5,105).
The rate of getting stuck in repetitive reasoning loops falls from 22.9% to 1.4% on IFBench. Runs hitting the 32k-token limit drop from 28.3% to 0.2%.
For on-device deployment, this matters as much as the parameter count. Shorter, more decisive reasoning means lower latency on constrained hardware—exactly the target use case.
Safety: Red-Teaming and Prompt Injection
In automated red-teaming across six harm categories, BotPsy produces unsafe responses 21% of the time versus 36% for its base model. That's the lowest rate of all models up to 1.2B parameters tested (LFM2.5-1.2B-Think: 35%; MiniCPM5-1B-Think: 47%).
Under prompt-injection attacks—critical for agents that read untrusted tool outputs—the unsafe rate drops from 46% to 31%.
These aren't enterprise-grade safety numbers, but for a sub-billion model they're respectable. Tether explicitly frames BotPsy as "not intended for high-stakes use" and emphasizes it's best suited to low-to-medium difficulty tasks.
The Broader Vision: P2P, Privacy, QVAC Fabric
Tether sketches a roadmap where BotPsy integrates with Hypersearch for decentralized P2P queries, bypassing corporate trackers. QVAC Fabric local fine-tuning will adapt to individual workflows and tool environments. Future iterations will delegate heavy computation to home PCs or trusted P2P nodes during offline-online transitions, while keeping all personal data, credentials, and decisions strictly on local hardware.
The vision is a "fully autonomous, highly capable personal AI ecosystem that keeps users entirely in control of their digital lives."
This is aspirational product language, but the technical foundation is real. BotPsy ships under Apache 2.0 in three precisions (bf16, Q8, Q4), as Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp.
Why This Matters
The industry conversation around agents has been dominated by frontier models—reasoning tokens, long-context windows, massive parameter counts. BotPsy makes a different bet: that most routine agentic work doesn't need frontier intelligence, and that the privacy, cost, and reliability wins from on-device deployment are worth optimizing for.
The benchmark results suggest this bet is viable. An 800M model that beats 2B rivals on tool use, that knows when not to call functions, that follows instructions reliably, and that runs efficiently on-device is a different kind of capability unlock than scaling to 70B or 400B.
It's also a reminder that the "best" model depends entirely on the deployment context. If your agent needs to reason through novel multi-step plans or synthesize information across vast contexts, you need frontier scale. But if you're automating calendar management, parsing email, or controlling IoT devices, BotPsy's performance envelope might be exactly right—and you avoid the operational cost, latency, and privacy surface of a cloud API.
The real test will be production reliability: how often does the 800M model fumble tasks that a 7B or 70B model would handle cleanly? How much prompt engineering and workflow constraint is required to keep it on-rails? And how much does the efficiency advantage erode once you account for retry logic and error handling?
Those questions won't be answered by benchmarks. But as a technical milestone—a sub-billion agentic model that legitimately competes with 2B rivals—BotPsy is worth paying attention to.