# BotPsy 0.8B: Tether's Open On-Device Agent Model for Tool Calling

> Source: <https://explainx.ai/blog/botpsy-0-8b-on-device-agent-qvac-apache-tool-use-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-11 00:00:00+00:00

**On November 11, 2026 Tether AI Research published [BotPsy](https://huggingface.co/blog/qvac/botpsy?ref=explainx), a **0.8B-parameter** agent model built on **Qwen3.5-0.8B** that targets **function calling** and **multi-step tool workflows** on **phones and laptops**—with **Apache 2.0** weights for research and education.** The [Hugging Face post](https://huggingface.co/blog/qvac/botpsy?ref=explainx) argues many routines do not need frontier cloud models, and that **sub-billion** agents can be **cheaper**, **more private**, and **less hostage to API policy changes**.

If you are shipping **local agents**—calendar helpers, device automations, or personal research bots—BotPsy is one of the few **sub-1B** releases that publishes **six agent benchmarks** and a **full training recipe**, not only a merge on top of a base chat model.

## TL;DR: BotPsy at a glance

| Question | Answer (per Tether's Nov 11, 2026 post) | 
|---|---|
| Base model | **Qwen3.5-0.8B** | 
| Size class | **0.8B** parameters | 
| License | **Apache 2.0** (research and education) | 
| Formats | **Safetensors** +**GGUF** (bf16, Q8, Q4) for QVAC /**llama.cpp** | 
| Headline claim | **~2x** agentic scores vs base on BFCL v4, ACEBench, τ-bench | 
| Vs 2B sibling | Beats **Qwen3.5-2B-Think** on**5/6** tool benchmarks in their table | 
| Safety | **21%** unsafe rate in their red-team harness vs**36%** for base | 
| Intended use | **Low–medium** difficulty personal assistant tasks—not high stakes | 

Numbers are **author-reported** unless you reproduce the eval harness yourself.

## Why another sub-billion agent model matters

Cloud agents are easy to prototype and expensive to operate at scale. A personal assistant that **searches**, **reads mail**, or **calls home APIs** can burn **millions of tokens** if every step hits a frontier endpoint. Small models also reduce **vendor lock-in**: the Hugging Face post explicitly lists **deprecation risk** and **limited customization** as reasons to host locally.

Tether already ships on-device speech with [TranscribePsy](https://explainx.ai/blog/tether-transcribepsy-250m-on-device-speech-recognition-2026). BotPsy extends the same **QVAC** story from **audio** to **tool-use agents**, sitting beside other **local** stacks covered in [Perplexity Computer running locally on a Mac](https://explainx.ai/blog/perplexity-computer-fully-local-mac-qwen-3-6-pplx-pro-36gb-offline-what-stays-on-device-2026) and [Liquid on-device reasoning models](https://explainx.ai/blog/liquid-ai-lfm2-5-24b-a2b-open-weight-on-device-reasoning-agent-model-2026).

Local on-device agent language model running tool calls without cloud API dependency

## Benchmarks: what BotPsy claims to beat

### Tool-use and agent suites

Tether compares **BotPsy-0.8B** against **Qwen3.5-0.8B-Think** (base), **Qwen3.5-2B-Think**, **LFM2.5-1.2B-Think**, and **MiniCPM5-1B-Think** on six benchmarks: **BFCL v4**, **ACEBench**, **τ-bench**, **Claw-Eval**, **BrowseComp-Plus**, and **ToolSandbox**.

Reported highlights from the blog:

| Benchmark | BotPsy-0.8B | Qwen3.5-0.8B 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 | 
| ToolSandbox | **64.6** | 30.5 | 41.2 | 
| Claw-Eval | 46.5 | 39.3 | **48.9** | 
| BrowseComp-Plus | **4.5** | 1.9 | 2.8 | 

The blog states BotPsy **roughly doubles** base tool-use capability and leads the **≤1.2B** class on **five of six** suites.

BotPsy 0.8B on-device agent model tool-use benchmark scores versus sub-1.2B rivals
*Image: [Tether AI Research / Hugging Face](https://huggingface.co/blog/qvac/botpsy?ref=explainx)*

### Multi-turn customer-service style tasks

On **τ²-bench telecom**, BotPsy reports **82.5** vs **69.3** for Qwen3.5-2B-Think. On **BFCL v4 multi-turn**, it rises from **7.5** (base) to **46.1**, which the authors contrast with **23.8** for the 2B model.

That matters because **multi-turn** tool loops are where tiny models usually collapse—either they **forget policy** or they **call tools with drifting arguments**. The results rhyme with why enterprises benchmark assistants on [τ²-style benches](https://explainx.ai/blog/raindrop-reasoning-agent-memory-tau2-bench-67-percent-2026) instead of single-shot trivia.

### Knowing when **not** to call tools

A capable agent must **decline** when no tool applies. On **BFCL v4 hallucination tests**, BotPsy scores **86.1**, the highest among models tested in the post. On **ACEBench special cases** (incomplete requests, wrong parameters), it scores **51.3** vs **13.7** for the base model.

That behavior is underrated in hobby agent demos but decisive in production—see [eBay blocking third-party shopping agents](https://explainx.ai/blog/ebay-blocks-third-party-ai-agents-browsing-shopping-2026) for how **uncontrolled tool loops** create platform risk.

### Knowledge vs tool specialization

Tool training did not erase knowledge entirely: **GPQA Diamond** moves from **33.2** to **38.1** and **MMLU-Pro** from **41.6** to **48.9** in the authors' tables, best among **≤1.2B** models tested. The blog still warns that **knowledge scales with parameters** and BotPsy should **call tools** for facts rather than act as a standalone knowledge model—similar philosophy to [Hugging Face System One decision APIs](https://explainx.ai/blog/hugging-face-decision-models-systemone-api-2026) that push **structured decisions** instead of long free text.

## Training recipe: four stages, three feedback types

BotPsy is not a single SFT pass. The [Hugging Face blog](https://huggingface.co/blog/qvac/botpsy?ref=explainx) describes:

1. **Supervised fine-tuning** on multi-turn tool trajectories, parallel and multi-step calls, and**no-tool** cases.
2. **Multi-teacher on-policy distillation** from three**Qwen3.5-4B** specialists (tool use, reasoning, instruction following) with per-token reverse KL.
3. **RL round 1** with**DAPO** and fine-grained tool-call rewards inspired by ToolRL-style partial credit.
4. **RL round 2** mixing math, code, knowledge MCQs, instruction following, structured outputs, and simulated workplace tool use.

The post includes stage-by-stage capability curves showing **SFT boosts tools but hurts knowledge**, then **distillation recovers** general skills before RL adds further gains—useful reading if you are building your own [agent skills](https://explainx.ai/blog/what-are-agent-skills-complete-guide) distillation pipeline.

## Safety, efficiency, and deployment notes

### Red-teaming numbers

Automated red-teaming across **six harm categories** reports **21%** unsafe responses for BotPsy vs **36%** for its base and **35–47%** for some 1B-class rivals in the same table. Under **prompt-injection** attacks tied to untrusted tool output, unsafe rate falls from **46%** (base) to **31%**.

Those rates are still **too high** for unsupervised customer support. The blog's own **Intended Use** section says BotPsy is **not** for high-stakes use.

### Reasoning token efficiency

On questions all five models answer correctly, BotPsy uses about **3× fewer reasoning tokens** than base on GPQA Diamond and **2.7× fewer** on MMLU-Pro, with loop rates collapsing on IFBench in their charts. For on-device latency, **shorter chains** matter as much as raw accuracy.

### How to try BotPsy

The post links **Hugging Face safetensors** and **GGUF** quantizations for **QVAC SDK** and [**llama.cpp**](https://github.com/ggml-org/llama.cpp?ref=explainx). Typical local stack:

1. Pull weights from the model card linked in the blog.
2. Pair with a **host app** that exposes tools (calendar, search, files) with explicit schemas.
3. Add **policy prompts** and**human confirmation** for irreversible actions—the model will not supply enterprise governance for you.

Compare setup friction with [Janestreet's agent-skills tutor repo](https://explainx.ai/blog/janestreet-agent-skills-socratic-tutor-github-2026) if you want a skills-first harness instead of a monolithic assistant binary.

## What Tether says comes next

The vision section mentions **Hypersearch** P2P queries, **QVAC Fabric** local fine-tuning, delegating heavy compute to home PCs, and **agent-to-agent** messaging on trusted LANs—all **on-device control** themes. Treat that as **roadmap prose**, not shipped product.

## Hardware and runtime expectations

Tether positions BotPsy for **phone-class** deployment via **GGUF** quantizations. Realistic planning:

| Precision | Likely role | Caveat | 
|---|---|---|
| **bf16** | Laptop prototyping | Higher RAM, best tool-call fidelity in author tests | 
| **Q8** | Daily-driver phone experiments | Watch thermals on long multi-turn loops | 
| **Q4** | Offline demos | Expect JSON/schema errors to rise; revalidate tools | 

You still need a **host process** implementing schemas—same lesson as [Expo Exact contract runtime](https://explainx.ai/blog/expo-exact-contract-agent-native-cross-platform-runtime-2026) posts about **typed tool surfaces**. BotPsy does not ship a calendar, browser, or payments stack.

On Apple silicon, compare latency against [Liquid's on-device reasoning line](https://explainx.ai/blog/liquid-ai-lfm2-5-24b-a2b-open-weight-on-device-reasoning-agent-model-2026): BotPsy is **smaller** but **more tool-specialized**.

## Benchmarks readers should rerun before betting products

Tether lists six suites—each measures a different failure mode:

- **BFCL v4** — schema adherence and multi-turn calls.
- **ACEBench** — messy user phrasing and special cases.
- **τ-bench / τ²-bench** — long customer-service dialogues with policy.
- **Claw-Eval** — claw-style tool environments where 2B still wins slightly.
- **BrowseComp-Plus** — web research (everyone scores near zero).
- **ToolSandbox** — simulated app integrations.

Copy the authors' lesson: **optimize the benchmark that matches your app group**—commerce vs device control vs messaging—not the unweighted mean. Figure 7 in the blog breaks scores by **app group** for simple vs complex scenarios; mobile assistants heavy on **device settings** should study that slice before trusting the headline average.

## Open weights and licensing nuance

Apache **2.0** for research and education is broader than many "open" models that block distillation or commercial fine-tunes. Still read Tether's **Intended Use** guardrails: they flag **high-stakes** automation and document **21%** unsafe generations in their harness.

If you bundle BotPsy into a consumer app, your privacy policy must state **on-device** inference clearly—regulators already scrutinize **default cloud** assistants. Pair with [Hugging Face's open-source security tradeoff essay](https://explainx.ai/blog/hugging-face-open-source-ai-security-tradeoff-solaiman-2026) when security asks why you chose local weights.

## What people are asking

| Question | Honest answer | 
|---|---|
| Is BotPsy better than Claude or GPT for agents? | Not on hard tasks. It targets **low–medium** personal workflows at**sub-billion** cost. | 
| Can I fine-tune BotPsy on my mail? | Apache weights allow research fine-tunes; check the license scope and your privacy policy. | 
| Does BotPsy include tools? | **No.** It is weights plus eval claims; you build the tool host. | 
| How does it relate to Qwen3.5? | It **is**[Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B?ref=explainx) plus Tether's four-stage post-training. | 

## Limitations you should plan for

- **BrowseComp-Plus** scores are**single digits** for everyone—web research is still hard at this scale.
- **Lower-resource languages** lag on instruction following per the blog's Figure 2 discussion.
- **Banking-style τ³** tasks stay near zero for all models tested—do not point BotPsy at treasury workflows.
- Benchmarks are **internal reproductions** ; wait for community runs before betting a product roadmap on the table.

## Related reading on explainx.ai

*Model names, scores, and release details follow Tether's November 11, 2026 Hugging Face post.*