A 0.8B personal assistant that beats models 2.5 times its size at tool use: meet BotPsy ๐Ÿ“ฑ

Most agent news is about giant models in giant data centers. That's why a post on the Hugging Face blog this morning caught my eye. Tether AI Research released BotPsy, a 0.8B-parameter model built for function calling and agentic workflows, small enough to run on a phone or a modest home server. The weights are already on the Hub under Apache 2.0, in bf16, Q8 and Q4, as safetensors and as GGUF files for llama.cpp.

What it can do

BotPsy is a fine-tuned Qwen3.5-0.8B, aimed at everyday assistant work: calendar, email, messaging, files, device settings, travel, shopping and simple web search. The interesting part is how far a sub-billion model gets on agent benchmarks. Compared with its base, BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and ฯ„-bench from 23.2 to 51.5. On five of six agentic benchmarks it also beats Qwen3.5-2B-Think, a model 2.5 times its size. Multi-turn work shows the biggest jump: BFCL v4 multi-turn goes from 7.5 to 46.1, and on ฯ„²-bench telecom, where the agent has to guide a user through troubleshooting over many turns, it scores 82.5.

My favourite number is a quiet one. On BFCL's hallucination tests, which check that the model declines when no suitable tool exists, BotPsy scores 86.1, the best of all models tested. For an agent that touches your inbox, knowing when not to act is half the job.

The training recipe

This is the part I'll be studying. Training happens in four stages, each starting from the best checkpoint of the previous one:

  1. Supervised fine-tuning on tool-use conversations, including cases where no tool should be called.
  2. Multi-teacher on-policy distillation. The student generates its own answers, and one of three specialist Qwen3.5-4B teachers (tool use, reasoning, instruction following) gives feedback on every token via reverse KL. This stage mainly recovers the knowledge and instruction following that SFT had damaged.
  3. Two rounds of RL with DAPO, using rule-based rewards that give partial credit for matching tool names, argument names and argument values.

The result is not just better tool use. IFEval climbs from 59.7 to 90.8, and the model reasons more efficiently: about 3 times fewer reasoning tokens per correct answer on GPQA Diamond, and runs that get stuck in reasoning loops on IFBench drop from 22.9% to 1.4%. On a small device, shorter reasoning means faster answers and less battery.

Honest limits

The team is refreshingly clear about weak spots. Deep research stays hard (BrowseComp-Plus is only 4.5), knowledge still scales with size, and instruction following fades as instructions pile up over turns. It was trained mainly in English, and Hindi and Russian lag far behind. They say plainly that it makes mistakes and shouldn't be used for high-stakes decisions involving money, health or sensitive personal data. I appreciate that tone a lot.

Why this makes me happy

For me this is the direction that matters: private, cheap agents that live on your own hardware and don't break when a cloud provider changes its terms. A 0.8B model that knows when to call a tool, and when to step back, is exactly the kind of building block I'd love to try in a homelab setup or a small learning assistant. Download it, run it in llama.cpp, and see what a tiny agent can do. ๐Ÿ™‚

Sources

Reacties

Een reactie posten

Populaire posts van deze blog

Meta just shared the full recipe for a 30B agent model you can run at home ๐Ÿง 

When Meta released Muse Glimmer in August, the pitch was simple: an open-weight, Apache 2.0 model built for AI agents that run on your own Mac or PC. What was missing was the "how". On 19 October, Meta Superintelligence Labs published the full technical report, and for anyone who cares about distillation and local agents, it is one of the most useful reads in a while. What the model is Muse Glimmer is a dense model of about 29.6 billion parameters, including a 1.8B perception encoder, so it reads text and images and writes text. It has a 128K+ context window and was trained on data from more than a hundred languages. The design goal was strict: everything has to fit in 32 GB of GPU memory, and a smaller 4-bit variant targets 24 GB cards. The parts I found most interesting Dense beat mixture-of-experts here. That surprised me. The team trained scaling ladders for a dense model and two MoE designs with the same memory footprint. Under a fixed memory ceiling and their compu...
Meer lezen

SpaceX just asked to fly Starship hundreds of times a year ๐Ÿš€

Last week the U.S. Federal Aviation Administration took a quiet but important step: it opened a formal review of SpaceX's request to launch Starship from Starbase in South Texas up to 456 times per year . Today the site is cleared for only 25 flights a year. That jump is the clearest signal yet that Starship is meant to become routine hardware, not a rare test rocket. What the FAA is reviewing On 14 October 2026 the FAA published a Notice of Intent to prepare a Programmatic Environmental Impact Statement (PEIS) for SpaceX's Starbase plans. A PEIS looks at a whole program of operations rather than a single launch, which is exactly what you need if you want a reusable rocket to fly the way airlines fly planes. SpaceX's proposal comes in two phases, according to the Federal Register notice: Phase 1: up to 100 launches and 200 landings per year (100 Ship, 100 Super Heavy), any day or time of week. Phase 2: up to 456 launches and 912 landings per year at Starbase. La...
Meer lezen

AI that can actually "see" a crystal ๐Ÿ”ฌ

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen