GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

Microsoft Fara1.5 Hits 63% on Mind2Web and 86.6% on WebVoyager With a 9B Open-Weight Browser Agent

Microsoft Research shipped Fara1.5 on May 21 — a family of three computer-use agent models for browser tasks — alongside MagenticBrain, an orchestration model, and MagenticLite, the next-generation Magentic-UI application that combines them into a local agent stack. The release is Microsoft’s clearest statement yet that useful AI agents do not require frontier-scale compute.

Benchmark Numbers

On Online-Mind2Web, which tests autonomous completion of 300 tasks across 136 popular websites, Fara1.5-9B scores 63% task success rate — nearly double its predecessor Fara-7B and more than 14 points ahead of GUI-Owl-1.5-8B (49%), the prior best in class for comparable model sizes.

On WebVoyager, a real-web navigation benchmark, Fara1.5-9B hits 86.6% against a range of 73.5–80.2% for similarly sized models.

The full family spans three sizes:

ModelParamsTarget hardware
Fara1.5-4B4BEdge / on-device
Fara1.5-9B9BModest GPU / laptop
Fara1.5-27B27BSingle consumer GPU

All three are being released as open weights.

What MagenticLite Does

MagenticLite is the application layer. It combines Fara1.5 (browser computer-use) with MagenticBrain (reasoning, delegation, and terminal access) into a single workflow that spans both the browser and local filesystem. The system is a direct successor to Magentic-UI.

The design goal is a capable agent that keeps data on the user’s machine — no cloud dependency required for basic operation. That positions it against products like Perplexity’s Personal Computer and OpenAI’s Codex Desktop, both of which require cloud routing.

Why It Matters

The dominant narrative in AI agents has been that performance scales with model size and cloud access. Fara1.5 is evidence that specialized training on browser interaction tasks — rather than raw scale — drives most of the practical gain.

Fara1.5-9B outperforms GPT-4V-based systems on the Mind2Web benchmark at a fraction of the compute footprint. The 4B model is intended to run fully on-device, which has implications for enterprise deployments with data residency requirements and consumer hardware as a viable agent host.

Microsoft Research framed the project as exploring “how far agentic performance can be pushed with smaller models, codesigned tools, and an optimized execution harness.” The Fara1.5 numbers suggest the answer is: farther than anyone expected eighteen months ago.

Context

This release comes three days after Arena added Gemini 3.5 Flash to its text and code leaderboards and two weeks after vix’s Claude Opus 4.7 scaffold hit 90.2% on Terminal-Bench 2.0 — the current frontier for agentic coding. Fara1.5 operates in a different domain (browser computer-use rather than coding agents), but the performance trajectory is consistent: specialized small models are closing the gap with general large models on specific task categories faster than the field expected.