Last night’s AliceAI-Foundation-80B note was a from-scratch base: 80B / 3B, KDA:GA 3:1, WikiWebFacts / HardMultiQA. Yesterday’s DeepSeek-V4.1-Flash and MiMo-V2.6-Pro pieces were API KV math and an open-weights AA 46 price floor. This morning is a different purchase: a generalist computer-use agent that clicks, types, runs code, and calls MCP or business APIs from one model — with published cost per task, not just pass rates. On September 28, 2026, H Company shipped Holo4.

Primary sources: H Company newsroom, Hugging Face blog, Holo4-27B card, Holo4-35B-A3B card, and the open trajectories dataset. Benchmark rows below are H Company’s harness (mean over 2–4 runs; single run on OSWorld 2.0 and ALE-CLI) unless marked as public leaderboard / other-provider charts. Treat every $/task as their token accounting at H Models API rates (or Alibaba list for Qwen foils) until you reproduce it.

Spec sheet (named numbers)

Holo4-27B Holo4-35B-A3B
HF checkpoint Hcompany/Holo4-27B Hcompany/Holo4-35B-A3B
Base architecture Qwen3.8 dense Qwen3.5 MoE
Size 27B dense 35B total / 3B active
Context (config) 262,144 262,144
Model type VLM (computer use) VLM (computer use)
Interfaces GUI + code + tool / MCP / API same
Weight license CC BY-NC 4.0 Apache 2.0
API list (USD / 1M tok) $0.40 in / $0.04 cached / $3.00 out $0.30 / $0.03 / $2.00
Formats on HF BF16, FP8, NVFP4, Q4 GGUF BF16, FP8, NVFP4, Q4 GGUF
Harness hai-agents (screenshots → actions) same

Sibling in the same drop: Holotron4 Nano (Nemotron 3 Nano Omni + same recipe). Newsroom deltas vs that base (absolute pp): OSWorld CUA GUI 21.0 → 76.3, AutomationBench MCP 19.4 → 35.6 — a transfer check, not the fleet SKU.

What actually changed for people who run agent fleets

The durable product change is one model across GUI, code, MCP, and API with a published score × $/task table and open evaluation trajectories.

Most agent stacks still pick a lane: a GUI specialist that dies without a screenshot, or a tool-calling model that stalls when the app has no API. Holo4’s pitch is that a real task mixes those modes — FreeCAD + shell + MCP in one session — and the same checkpoint should run on desktop, web, Android, code sandbox, and business APIs without swapping models.

Training stack (newsroom): Agentic Task Factory ~10,000 tasks (~4k web, ~3k MCP, ~3k desktop/OS), including hybrid GUI+MCP environments. SFT on 127B tokens (~¾ successful agentic trajectories: desktop 45%, web 14%, MCP/API 12%, mobile 3%). Then two asynchronous online RL LoRA experts (desktop/web vs terminal/MCP/API) merged at equal weight. They also rebuilt the harness (memory across hundreds of steps; shell on the desktop) and publish the OSWorld 2.0 climb from Aug 10 6.1% to release 61.7%. Trajectories behind the public scores are open — replay at trajectories.hcompany.ai or pull Hcompany/trajectories.

Vendor benches vs what you should not invent

H Company table (their harness; under each score: cost per task USD at H Models rates for Holo4; Qwen foils at Alibaba list from their token counts; frontier rows are public / other harnesses — do not treat as A/B):

Benchmark Holo4-27B Holo4-35B-A3B Qwen3.8 27B Named frontier foil
OSWorld (computer tasks) 85.2% / $0.08 80.8% / $0.05 84.3% / $0.22 Fable 5 86.0%; Qwen3.8 Max 86.1%
OSWorld 2.0 (long workflows) 61.7% (succ. 41.5%) / $1.22 30.9% (12.3%) / $0.61 48.0% (19.4%) / $3.49 Opus 5.5 81.8% (48.7%) / $8.48; GPT-6 Astra 73.5% / $9.07
ALE-CLI (105-task Linux) 44.1% (19.4%) / $0.82 30.9% (13.5%) / $0.29 43.5% (19.0%) Opus 5.5 63.7% / $8.22; GPT-6 Astra 61.4% / $5.31
AutomationBench v1.0.6 public 45.4% / $0.05 34.5% / $0.02 40.3%* / $0.09 Opus 5 50.3% / $3.05; GPT-5.6 Sol 45.8% / $0.67
AndroidWorld 85.1% / $0.08 77.6% / $0.07 81.9% / $0.13 Fable 5 88.8%; Qwen3.8 Max 85.3%

* Qwen3.8 27B in H Company’s harness where no public score exists. AutomationBench footnote: 480/600 public tasks overlap their training-data split; held-out 120: Holo4-27B 49.3%, 35B-A3B 31.7% vs Qwen3.8 27B 40.3%. They will report the official private set when evaluated — do not paste public-set as private-set.

Agentic Task Factory held-out (not used in training): Web 80.2% / MCP 89.4% / Desktop 72.0% for 27B (35B-A3B: 70.8% / 85.4% / 64.8%).

Pattern for a routing ticket: 27B wins the hard long-workflow row (OSWorld 2.0 61.7%) at ~1/7 the Opus 5.5 $/task in their chart; 35B-A3B is the cheap Apache-2.0 MoE sibling that under-indexes on long desktop workflows but still posts AutomationBench 34.5% at $0.02/task. Neither claims an Artificial Analysis Intelligence Index row in this drop — do not invent one.

Where it sits next to AliceAI, Flash, MiMo, and closed agents

Purchase Product What it buys
From-scratch hybrid base + RU protocols AliceAI-Foundation-80B 80B/3B; KDA:GA 3:1; Apache-2.0 base
Model-side KV / activate API DeepSeek-V4.1-Flash 890 B/tok; 8B/16B; Flash rates
Open-weights AA price floor (instruct) MiMo-V2.6-Pro AA 46; $0.435/$0.87; 1.02T/42B
Computer-use $/task + open trajectories Holo4 OSWorld 2.0 61.7% @ $1.22; Apache 35B/3B; GUI+code+MCP
Default coding agents (OpenAI) GPT-6 Sol $2/$10; cache-steerable effort
Volume / extract (OpenAI) GPT-6 Luna $0.10/$0.50
Flagship / agentic coding (Anthropic) Claude Opus 5.5 Fable-class work at $4/$20
Flagship reasoning (OpenAI) GPT-6 Astra Highest OpenAI ceilings; $10/$50

Holo4 does not claim an AA ceiling. It argues that a mid-size VLM with published GUI+code+MCP cost tables and open traces is a better agent-fleet canary than guessing which chat flagship still works when the app only has a screen.

Caveats I would write into any Holo4 ticket

  1. Harness ≠ model. Scores are Holo4 inside hai-agents. Swap the loop and the number moves — reproduce before you retire a Sol / Opus path.
  2. License split is real. 35B-A3B is Apache 2.0; 27B is CC BY-NC 4.0. Self-host commercial fleets should start on the MoE card unless counsel clears NC.
  3. OSWorld ≠ OSWorld 2.0. Short-task 85.2% is not long-workflow 61.7%. Quote the row you mean.
  4. AutomationBench public set overlaps training data (480/600). Prefer held-out 120 (49.3% for 27B) or wait for private-set.
  5. No AA Index in this drop. Keep MiMo Pro on the open Intelligence Index floor.
  6. Not a rematch of Inferra / Leo X / Lessengers / Continuum. Those are memory and optics. This is an agent model + harness ticket.
  7. Frontier foils are mixed harnesses. Opus / Astra / Fable $/task rows are order-of-magnitude, not a controlled bake-off.

How I would evaluate if I ran computer-use or MCP fleets

  1. API canary on H Models at published rates — your $/task on a fixed OSWorld 2.0 / AutomationBench subset before any weight download.
  2. Self-host Apache 35B-A3B (BF16 → FP8/NVFP4 as memory allows) behind hai-agents or your loop; keep 27B as research/NC only.
  3. Replay Hcompany/trajectories on failing tasks to separate harness bugs from weights.
  4. A/B against Qwen3.8 27B on the same harness, and against Sol / Opus 5.5 on the hard unattended subset where $/task still loses to closed max-effort.
  5. Keep AliceAI for post-train substrate and MiMo / Flash for chat/API volume — Holo4 is the computer-use board.

This is a computer-use agent release with named OSWorld / AutomationBench $/task, dual licenses, and open trajectories — not a new closed flagship and not an AA rematch. Put Holo4 where you need GUI + code + MCP in one mid-size VLM with a readable cost table. Keep MiMo on the open ceiling/price row, DeepSeek-V4.1-Flash on cacheable API loops, AliceAI on from-scratch bases, and Astra / Opus 5.5 on hard unattended work until your harness says otherwise. Measure, then move traffic.

Create-only. Specs from H Company’s Sep 28 newsroom, the HF blog, the Holo4-27B / Holo4-35B-A3B cards, and the trajectories dataset. Update when the AutomationBench private set lands, when Artificial Analysis publishes an Index row, when DSpark drafter checkpoints ship, or when a named third-party harness reproduces OSWorld 2.0 against Sol / Opus 5.5.

Sources: H Company newsroom · Hugging Face blog · Holo4-27B card · Holo4-35B-A3B card · trajectories dataset.

Anpoo Sivanadi, Staff Software Engineer