Tether BotPsy brings 800M-parameter AI tool calling to local devices | Magica
Tether releases BotPsy for local AI agents, with phone performance still unproven
Editorial Team
••📖8 min read
The USDT issuer’s 800-million-parameter BotPsy beats a larger Qwen model on five of six company-reported agent benchmarks. Its local-AI ambitions still face missing handset measurements and explicit limits on sensitive uses.
BotPsy’s weights are free under Apache 2.0; developers must build the assistant and control its access to tools.
Tether reports gains over its base model and wins against a larger Qwen model on five of six agent benchmarks.
Compact downloads do not establish phone speed, battery costs or safe access to sensitive information.
Tether, the issuer of the USDT stablecoin, released QVAC BotPsy on November 11, an 800-million-parameter AI model for developers building assistants on personal devices. It outperforms an Alibaba Qwen model 2.5 times its size on five of six agent benchmarks in Tether’s tests. The release documents do not establish how fast it runs on a phone—and warn against the sensitive uses featured in its launch.
Tether’s conceptual BotPsy launch illustration. Applications provide the connections, permissions and controls that execute the model’s tool calls. Image: Tether. Source: Tether.
A local model, with applications still doing the work
BotPsy comes from Tether AI Research, the name used for Tether Data, S.A. de C.V. Its QVAC initiative develops local and decentralized AI. Tether announced the QVAC development platform in May 2025, already proposing agents that would run without centralized cloud infrastructure. That announcement also described WDK, Tether’s payment integration, as enabling agents to transact in Bitcoin and USDT. The platform ambition therefore extends to payments, although the BotPsy release does not establish payment adoption or revenue.
BotPsy adds action selection to Tether’s existing range of vision, speech, health and translation models. Built on Qwen3.5-0.8B, an open model from Alibaba, it selects an application’s functions and supplies their arguments—for example, choosing a calendar function and specifying an appointment’s time. The application executes the call, returns the result and lets the model decide the next step.
That division matters. Tether’s announcement says applications provide service connections, permissions and execution controls. Its explicitly says there is no BotPsy app: developers must build the model into their own assistants.
Palantir’s co-founder and CEO says he could back a Democrat in 2028, while excluding a Democratic Socialists of America member. His Democratic history complicates talk of a political conversion; the more immediate question is how his government software company will answer congressional scrutiny.
Local processing can reduce the personal context sent to a model provider and avoid a hosted model’s per-request bill. It does not make every operation offline: search and booking tools can still require a connection. Nor does it remove the need to control what an assistant may read, change or send. In Tether’s ACEBench example, BotPsy recovered from a failed send by enabling Wi-Fi and deleting a message to free inbox space. That illustrates why permission to complete a task also requires limits on the intermediate actions it can take.
Company-reported benchmark scores from Tether AI Research. Metrics differ across tests; these results do not measure phone performance. Chart: Magica. Source: QVAC BotPsy.
The benchmark gains have a defined scope
Tether’s model card compares five models in thinking mode using a shared evaluation harness. BotPsy was served in BF16 precision through vLLM; benchmarks using external APIs ran on November 3, 2026. These are developer-reported evaluations, rather than measurements of everyday phone use.
Company-reported benchmark score
BotPsy 0.8B
Qwen3.5-0.8B-Think base
Qwen3.5-2B-Think
BFCL v4
44.0
23.4
37.7
ACEBench
58.7
30.1
55.4
τ-bench
51.5
23.2
46.5
Claw-Eval
46.5
39.3
48.9
BrowseComp-Plus
4.5
1.9
2.8
ToolSandbox
64.6
30.5
41.2
All scores are percentages, with higher better, but the measures differ. BFCL uses its official weighted score; ACEBench measures accuracy; the reported τ-bench combines retail, telecom and banking subsets, excluding airline. Claw-Eval multiplies safety by a weighted combination of completion and robustness; ToolSandbox measures mean similarity over eight settings. They should not be read as one common real-world success rate.
BotPsy beats the two-billion-parameter Qwen comparator on five rows, while losing on Claw-Eval. It also leads the tested LFM2.5-1.2B-Think and MiniCPM5-1B-Think models across the six tests. This supports a claim about the tested small-model field, rather than AI agents generally.
On BFCL’s multi-turn component, BotPsy scores 46.1, against 7.5 for its base and 23.8 for the larger Qwen. It also scores 86.1 on BFCL’s test of declining to call a tool when none is appropriate, the highest among the models tested.
The weaknesses remain consequential. A leading BrowseComp-Plus score of 4.5 is still low in absolute terms; no tested model scores above 5 on the banking subset. The larger Qwen model also leads BotPsy on general-knowledge tests. Tether’s regrouping of multi-step tasks by application yields a BotPsy average across seven app groups of 50.3 out of 100, against 22.1 for its base model and 42.5 for the larger Qwen. Those groups use benchmark metadata where available and an AI model to classify remaining scenarios; they are another view of test tasks, rather than observed app usage.
Tether post-trained the Qwen base in four stages: supervised fine-tuning, multi-teacher on-policy distillation and two rounds of reinforcement learning. The weights are released under Apache 2.0, but training code and training data are not included.
Small local action models also predate this release. Google released FunctionGemma in December 2025, specializing its 270-million-parameter Gemma 3 model for function calling and further customization. Google reported accuracy on its held-out Mobile Actions evaluation rising from 58% to 85% after fine-tuning. It also offered a fine-tuning workflow for commands such as calendar updates and a phone-based game demo. These are concrete alternatives for developers, but neither the different evaluation nor the smaller parameter count establishes a ranking against BotPsy: FunctionGemma is absent from Tether’s comparison.
Company-reported GGUF file sizes from Tether AI Research. Download size does not establish working-memory needs, phone speed or battery consumption. Chart: Magica. Source: QVAC BotPsy GGUFs.
Smaller files are only part of the deployment economics
Tether’s GGUF release, a model-file format supported by the llama.cpp inference engine and Tether’s QVAC development kit, includes files of about 1.52 GB in 16-bit BF16, 0.81 GB in 8-bit Q8_0 and 0.53 GB in 4-bit Q4_K_M. Those are model-file sizes, not total working-memory requirements.
The company recommends Q8_0, which stayed within about one percentage point of BF16 GGUF on nine of ten evaluations and scored 1.9 points higher on IFBench. The smallest version lost more on knowledge tests: MMLU-Pro fell from 46.0 for BF16 GGUF to 42.5, while GPQA Diamond fell from 39.2 to 35.3. ToolSandbox was not run on the GGUF files.
The GGUF evaluations used NVIDIA H100 GPUs on November 6–7. Changes in engine, run date and repeat count complicate comparisons with the original vLLM results. They do not establish mobile latency, energy consumption or sustained performance.
Tether’s technical guide also reports fewer reasoning tokens: 3,690 per correct GPQA Diamond science-test answer versus 11,563 for the base model, measured on questions all compared models answered correctly. That narrower efficiency result is not a phone-speed measurement. The release documents provide no comparable handset latency, battery or total deployment-cost figures. Developers also need to enable thinking mode explicitly: the supplied chat template defaults to non-thinking, while the reported evaluations used reasoning before an answer.
Company-reported unsafe-response rates from Tether AI Research. Lower is better; unsafe responses remain in both attack evaluations. Chart: Magica. Source: QVAC.
The next test is controlled execution on real devices
The safety results reinforce the gap between local control and dependable action. In its automated tests across six kinds of harm, Tether reports unsafe responses in 21% of attacks, versus 36% for its base model. In prompt-injection tests—where hostile instructions can arrive inside material an agent reads—the unsafe rate fell from 46% to 31%. The company recommends a separate input and output filter for applications with strict safety requirements.
Its model card goes further: BotPsy is intended to support research and education and is not intended for safety-critical or high-stakes automated decisions, or applications involving money, personal information, health and other sensitive topics. That restriction sits uneasily beside the calendar, inbox and commerce examples used to explain its potential. English is the recommended language, and low- to medium-difficulty tasks are its stated target.
Tether’s longer-term plan would make a phone the interface while a home computer or trusted peer-to-peer nodes handle heavier work. Encrypted peer-to-peer search and personal fine-tuning through QVAC Fabric are described as future capabilities. They should not be counted as features already delivered by the model download.
The next decision is how much authority an application should give this model. Reproducible tests on named phones would establish speed and resource use; tests of permissions, error recovery and prompt injection would help establish which actions it can safely attempt. Until those questions are resolved, the weights support experimentation, while Tether’s own limitations argue against handing BotPsy control of sensitive accounts.
ByteDance’s video model is reaching creators through multiple platforms with 30-second 4K output. Posted prices and draft workflows show why the choice of distributor still matters.
Rox’s $1.3 billion valuation backs AI agents for large enterprise sales teams. Its growth and cost claims face a crowded market, while integration with existing systems leaves the case for replacing Salesforce unresolved.
Google’s Gemini Decision 1 scores predefined choices across text, images, audio, video and PDFs for $0.10 per million input tokens. Its shared-input billing targets recurring workflow costs, but cheaper decision models, rival caching and low-cost generative alternatives complicate the savings claim.
Treasury and other agencies reportedly want Singapore cloud provider Megaspeed on a U.S. trade blacklist. Commerce denies a reported cabinet confrontation over the delay, while Nvidia says it has blocked shipments since June 2025.
United is restricting unauthorized AI agents from searching, booking and managing travel through its platforms and intermediaries. The airline cites security and safety risks as it reserves control over future integrations.
A controlled study found that pharmaceutical ads shifted AI recommendations toward promoted drugs without reducing overall clinical accuracy. The results expose a gap in accuracy tests, but do not establish effects on prescriptions or patient outcomes.
Former AWS computing chief Dave Brown will succeed Santosh Janardhan in early 2027 as Meta weighs its next capacity budget after a quarter of negative free cash flow.
Google is rolling out its Gemini app on Apple CarPlay, bringing voice conversations linked to Gmail and Google Calendar. It joins existing in-car AI options, while some personalized features remain restricted by age and region.
Google Labs’ Tabletop lets groups generate and use shared itineraries, polls and other interactive tools. The US adult-only experiment builds on an idea already offered by shareable AI app builders, leaving its practical advantage to be tested.
Anthropic’s IPO has reportedly attracted orders multiple times the available shares. Its planned $3 billion–$3.5 billion Japanese allocation would exceed Japan’s reported domestic IPO proceeds so far this year, but indications of interest are not completed sales.
Google warned that charging for news links would force it to stop distributing Indonesian news, while Meta and TikTok challenged other parts of the copyright bill. Existing Indonesian payment mechanisms and overseas deals complicate the platforms’ warnings about the cost to publishers.
Utrecht is investigating restrictions on camera glasses in municipal buildings, with findings expected in early 2027. The city must establish whether a ban is lawful, enforceable and compatible with assistive uses.
Research commissioned by cloud-storage supplier Wasabi estimates £3.8 billion in annual UK infrastructure overspend attributable to hard-to-forecast provider charges, including £1.2 billion for AI. Businesses report hiring and investment trade-offs, but recovering that money and turning it into jobs remain conditional.
Google DeepMind’s AlphaGenome 2 improves cell-specific DNA predictions and accepts new single-cell data without fine-tuning. Company benchmarks show gains at high confidence, but limited recall, complementary tools and greater computational demands leave experimental validation essential.
Coinbase is rolling out server-enforced trading limits and portfolio-loss triggers for AI agents connected to customer accounts. Product-category permissions remain forthcoming, and pausing new trades does not establish a ceiling on losses.
JetBrains IDEs 2026.3 bundle Air for running coding agents and reviewing their changes, while the standalone Air app is in maintenance mode. Existing provider subscriptions remain usable, but billing, security controls and execution options vary by integration.
Amazon’s shares and convertible notes could be worth $271.8 billion at the top of Anthropic’s proposed IPO range. Existing accounting gains, a 12% ownership cap and Anthropic’s computing commitments shape what that paper value means.
Nvidia chief Jensen Huang argued in May 2025 that US chip restrictions were strengthening Chinese rivals. Filings show a substantial commercial stake, but lost sales and market share alone cannot establish whether the controls failed their security purpose.
Reported demand has Anthropic’s bankers considering a price above its filed $70–$78 range. The combined financing could exceed $100 billion, but existing cash, uneven operating results and long-term computing commitments complicate the valuation test.
