What happened
Tether AI Research’s announcement, published on November 11, 2026, introduces BotPsy as a personal-assistant model designed for function calling and agentic workflows. The company says it is based on Qwen3.5-0.8B and is best suited to low- and medium-difficulty tasks. Its described domains include web search and research, travel and lifestyle, commerce and finance, business and customer operations, device and vehicle control, files and documents, calendars, and email and messaging.
The announcement reports that BotPsy outperformed the models of up to 1.2 billion parameters that Tether tested on all six listed tool-use and agent benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval and BrowseComp-Plus. It also reports higher scores than Qwen3.5-2B-Think, which it calls 2.5 times larger, on five of those six benchmarks. For example, BotPsy scored 44.0 versus 37.7 on BFCL v4, 58.7 versus 55.4 on ACEBench, and 51.5 versus 46.5 on τ-bench; Qwen3.5-2B-Think led on Claw-Eval, 48.9 to 46.5. Against BotPsy’s base model, the announcement lists gains on BFCL v4 from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5.
The authors describe a four-stage post-training process: supervised fine-tuning, multi-teacher on-policy distillation, and two reinforcement-learning rounds. The distillation stage uses three specialist Qwen3.5-4B teachers trained for tool use, general reasoning and instruction following. The reinforcement-learning stages use DAPO and rule- or verifier-computed rewards. The announcement also reports a 21% unsafe-response rate in automated red-teaming across six harm categories, compared with 36% for the base model; under prompt-injection attacks, it reports 31%, down from 46%.
The announcement says the weights are released under Apache 2.0 for research and educational use, in bf16, Q8 and Q4 formats as Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp. It cautions that the model makes mistakes and is not intended for high-stakes use. The company does not present BotPsy as a general replacement for larger models: its stated target is a range of routine tasks where local deployment may be appropriate.
What the archive adds
New Runtime’s earlier coverage described Meta’s Muse Glimmer 30B as an open-weight agentic model designed for consumer hardware (October 22, 2026 note). BotPsy is a new event in the same broader question of running agentic workloads on local hardware, but it differs in scale and reported emphasis: the announcement centers on a sub-billion-parameter model specialized for tool use.
Trend signal
This supports the pattern that agent economics may move toward the cost of completed work rather than model calls alone: agent economics moves to completed work. BotPsy’s announced size and tool-use scores make small-model task performance a relevant comparison point, although the release does not establish real-world cost per verified task. This does not yet validate the economic claim in production.
Why now
Tether AI Research frames BotPsy against the operational cost, access uncertainty and privacy trade-offs of cloud frontier models. The announcement positions on-device models as a way to keep some routine assistant tasks local while reserving larger systems for harder work. That is the company’s rationale, not evidence that BotPsy already delivers those benefits in deployed applications.
Why it matters
A small model that can call tools reliably could make some agent workflows easier to run on devices or private infrastructure, where teams may care about data control, connectivity or predictable resource use. BotPsy’s benchmark results give builders a reason to test whether specialization can substitute for model scale on narrow tasks. The most useful signal is not simply that a 0.8B model scores well, but that the release reports gains in multi-turn tool use and in declining when no suitable tool exists.
Those claims still need to survive evaluation against a team’s own APIs, policies, languages and failure costs. A lower parameter count does not by itself guarantee low latency, lower total cost, safe tool execution or acceptable quality. Teams should treat BotPsy as a candidate component for bounded work, with permission controls and human review appropriate to the consequences of its actions—not as an autonomous personal assistant ready for high-stakes use.
Limits and uncertainty
The performance figures in the announcement are reported by the model’s authors; they are not independent replications. The release’s summary does not establish how BotPsy performs on a particular team’s tools, hardware, languages or live workloads, and benchmark scores do not prove reliable completion in deployment. The authors explicitly warn that the model makes mistakes and is unsuitable for high-stakes use. Actual latency, memory requirements, energy use, and end-to-end cost on target devices require measurement in the intended environment. The reported red-teaming and prompt-injection percentages also do not establish safety against all attacks or tool configurations.
Skill to build now
Build a small local evaluation set from one routine workflow your team might delegate: include valid tool calls, missing or incorrect arguments, requests that should trigger no tool, and multi-turn changes to user intent. Run the same cases against the base model and BotPsy, then record task completion, unnecessary actions, refusals and resource use. This turns the announcement’s benchmark claims into a decision about your own workload.
Reader takeaway
BotPsy is a credible candidate for experiments with local, bounded tool use, not proof that a tiny model can replace a general-purpose agent. Its reported scores justify testing the model on real workflows, while its stated limitations make careful evaluation and constrained permissions essential.