Tether's QVAC BotPsy-0.8B Doubles Tool-Use Scores, Beats Models 2.5x Its Size
NielsRogge · x · 2026-11-15
Tether's QVAC has released QVAC BotPsy-0.8B, an open-weights agent model specialized from Alibaba's Qwen3.5-0.8B-Think for on-device tool use, roughly doubling its base scores and outperforming models up to 2.5x its size.
Benchmark gains across six agentic evaluations:
- BFCL v4: 23.4 → 44.0
- ACEBench: 30.1 → 58.7
- τ-bench: 23.2 → 51.5
- ToolSandbox: 30.5 → 64.6
- Also covers Claw-Eval and BrowseComp-Plus
The thesis: rather than scaling general-purpose models, developers can specialize small models for specific agentic tasks within the memory and compute limits of personal hardware. Weights are Apache 2.0 on Hugging Face, targeting personal assistants running on user-controlled devices.
More from coding & agent
- iceberg-mcp-server: an open MCP server to query and manipulate Apache Iceberg tables — dragonejt · 2026-11-15
- Kyvenza 5.0 turns Mac VMs into AI agent workspaces with MCP file, shell and mouse tools — ScholarBackground836 · 2026-11-15
- A robot band in your browser: physics-simulated instruments you can play via MIDI — Active_Cauliflower71 · 2026-11-15
- Claude writes 12,500 lines of Blender Python to make a 4K film, renders with 89 rented GPUs for $20 — john_lam · 2026-11-15
- Design engineering tip: show the first frame before motion content is ready — floguo · 2026-11-15
- FreeToken open-source engine runs 284B DeepSeek locally at 25 tok/s on RTX 5090 — aigclink · 2026-11-15