Is Orvena 4B state of the art?Pending benchmarks.Announced Nov 4, 2026; no independent results yet.
Orvena's open-weight on-device assistant model for the Orvena iPhone app: a LoRA fine-tune of Qwen3.5-4B (LoRA r=16, alpha 320 on all 248 linear layers, merged) that keeps the base architecture, including the vision encoder, and answers without a separate thinking step by default. Trained for phone tools (calendar, alarms, places, travel times, web) and MCP connections such as Notion, Todoist, Airtable and Gmail, in 30 languages. Apache-2.0 bf16 weights with a retrained MTP head, plus a 4-bit MLX build the app downloads. No hosted API stated.
The details so far
- Lab
- Orvena
- Announced
- Nov 4, 2026
- API
- Not yet · No hosted API stated; vLLM / Transformers / SGLang instructions for local serving
- Open weights
- Published
- Parameters
- 4B (Qwen3.5-4B backbone)
- Base model
- Qwen/Qwen3.5-4B (Alibaba); 24 Gated DeltaNet layers, 8 full-attention layers, vision encoder retained
- Training
- LoRA r=16, alpha 320 on all 248 language-model linear layers for one epoch, then merged; thinking off; MTP head retrained on the fine-tuned model's outputs
- Training data
- 3,397 simulated-iPhone scenarios (11,161 rounds) with Orvena's tool catalog and MCP tool lists from 12 services, plus no-tool/restraint sets and 600 IFEval-style instruction-following rows
- App
- Default model in Orvena 2.5 for iPhone; 4-bit MLX build is what the app downloads
- Languages
- 30 languages listed on the model card
- Checkpoints
- Orvena/Orvena-4B (bf16 + MTP) and Orvena/Orvena-4B-MLX-4bit on Hugging Face; technical report PDF on the card
- License
- Apache-2.0
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on the Orvena bench of 547 held-out scenarios, Orvena 4B passed 440 (80.4%), against 345 (63.1%) for Qwen3.5-4B and 104 (19.0%) for Apple's on-device model.
Lab claim: on multi-step tasks, 43 of 70 against 17 for the base model.
Lab claim: on an iPhone 17 Pro Max with the 4-bit MLX build, median time per task was 15.1 s (base thinking off: 16.3 s; thinking auto: 37.0 s).
Sources
REPOOrvena/Orvena-4B model card (Hugging Face)huggingface.co/Orvena/Orvena-4BREPOOrvena/Orvena-4B-MLX-4bit model card (Hugging Face)huggingface.co/Orvena/Orvena-4B-MLX-4bitLABOrvena announcement on Xx.com/orvenaAI/status/2118321172420960536 · Nov 5, 2026LABOrvena app siteorvena.appTracking continues until an independent board lists it.