H Company Holo4: computer-use agents — 27B CC BY-NC, 35B-A3B Apache-2.0

H Company announced Holo4 on 2026-09-28 (HF blog + newsroom): a series of agentic / computer-use vision-language models in two sizes—27B dense (Qwen3.8 base) and 35B-A3B Mixture of Experts (Qwen3.6 base; ~3B active naming). Both ship on the H Models API and as Hugging Face weights (HF blog, newsroom).

This is a Desk Bot models briefing. Companion Holotron4 Nano is out of this slug’s license lead—do not fold it into the Holo4 rights sentence.

HARD: license split (keep explicit)

Model	Weights license	Commercial self-host?
Holo4-27B	CC BY-NC 4.0 (non-commercial). Hub license:cc-by-nc-4.0. Card notes Qwen3.8-27B base is Apache-2.0; repo includes both licenses.	No for weights — use API for commercial 27B
Holo4-35B-A3B	Apache License 2.0. Hub license:apache-2.0. Base Qwen3.6-35B-A3B also Apache-2.0.	Yes under Apache-2.0

Do not collapse “Holo4 open-weights” into one rights story. API Terms ≠ weight redistribution rights (27B card, 35B-A3B card).

What it does

Same model family across interfaces: GUI click/type, code write/run in a sandbox, MCP, and APIs—on desktop, web, Android, code sandbox, and business APIs. Trained via supervised + RL on Agentic Task Factory tasks (~10k claimed). Harness: hai-agents (HF blog, newsroom).

Context: prefer “256K (config max 262,144 tokens)”—newsroom/API say 256K; cards list 262,144. Formats on HF: BF16, FP8, NVFP4, 4-bit GGUF (prefer BF16 over blog “FP16” shorthand). Trajectories open-sourced: viewer + Hcompany/trajectories (collection).

API pricing (as of Fact check)

Per models-api / newsroom (USD per 1M tokens):

• Holo4 27B: input $0.40 / cached $0.04 / output $3.00
• Holo4 35B-A3B: input $0.30 / cached $0.03 / output $2.00

Re-check models-api if citing dollars at ship time.

Soft: vendor harness benches (attribute)

All scores/costs below are H Company claims in H’s harness—not aitamer re-runs; harnesses/effort differ from frontier charts (newsroom, HF blog):

• OSWorld: 27B 85.2% @ $0.08/task; 35B-A3B 80.8% @ $0.05/task
• OSWorld 2.0: 27B 61.7% (41.5% success) @ $1.22; 35B-A3B 30.9% (12.3%) @ $0.61 — vs vendor-cited Opus 5.5 81.8% / GPT-6 Astra 73.5% (different harnesses)
• Footnotes: mean 2–4 runs except single-run OSWorld 2.0 / ALE-CLI; “Holo4 in our harness.”

Dense 27B ahead of 35B-A3B on those vendor OSWorld numbers is OK only as attributed vendor framing—not an independent ranking. No MarkTechPost as primary.

Who should care

Teams picking a downloadable computer-use stack should start at the HF blog and cards—keep the 27B CC BY-NC vs 35B-A3B Apache-2.0 split in every mention, soft-attribute every harness bench, and use API for commercial 27B rather than inventing weight rights.