Xenon Hunmin Computer-Use Release

Xenon open-sourced Hunmin VLM 397B on September 18, 2026, a computer-operating vision model built on Qwen 3.5-397B that locates on-screen elements and drives Linux and Windows tasks directly. It scored 75.6 on ScreenSpot Pro for second place of 48 listed models, 70.5 on OSWorld at plus 22.3 over base, and plus 9.1 on WindowsAgentArena. Post-training used eight Nvidia B200 GPUs with low-rank transfer plus supervised and reinforcement learning in FP8, and Xenon ships an FP8 variant halving storage alongside eval conditions and methods.

• What it does: answers plus acts, grounding buttons and fields to coordinates then executing clicks, typing, and navigation.
• Efficiency story: frontier-class grounding added to a 397B base with eight GPUs, not a datacenter, while keeping Korean-language performance.
• Openness: weights plus FP8 build plus eval harness released, feeding Xenon's OneAgent action-AI upgrades.

I run computer-use pilots at SaaSNext. Open weights with published evals change the buy. Here is the breakdown and the deploy.

Grounding Is the Benchmark That Matters

Computer-use fails at coordinates before it fails at reasoning. A model that plans perfectly but clicks three pixels left of Submit loops forever. ScreenSpot Pro measures exactly this: locate the element to operate across demanding screenshots. Hunmin at 75.6 for number two of 48 says its clicks land. OSWorld at 70.5 across 360 Linux tasks says those clicks compose into finished work at plus 22.3 over its own base. WindowsAgentArena plus 9.1 says the skill transfers across operating systems rather than overfitting one shell.

Treat vendor figures as dated evidence pending independent replication, the same discipline I use for DeepSWE vs Terminal-Bench vs SWE-Atlas selection. The shapes still inform: grounding gains transferred to task gains, and both came from post-training rather than a new pretrain. For buyers that means capability without a new foundation bet.

War Story 1: The Closed Clicker That Could Not Point

We piloted a closed computer-use API for invoice triage across a browser dashboard plus a desktop ERP. Planning traces read beautifully. Execution missed: date-picker clicks landed on adjacent weeks, file-dialog paths typed into search boxes, confirmation buttons hit at 80 percent with the rest landing on background. Three weeks of prompt tuning moved success from 41 to 47 percent. The vendor blamed our screenshots. We blamed the model. Both were guessing because grounding scores were unpublished.

Hunmin-style openness ends that argument. Published ScreenSpot numbers plus released eval conditions let buyers replay grounding before buying task success. I added a grounding gate to procurement this week: 200 screenshots from our apps, vendor model or ours, click-point accuracy required before any task trial. No grounding proof, no pilot.

• Before: closed API, 47 percent task success, zero grounding diagnostics, stalled rollout.
• After rule: grounding gate first, task trials second, prompt tuning last.

Do not tune prompts for a model that cannot point. Here is why coordinates precede reasoning.

Benchmark Breakdown With Dates

Figures below come from Xenon's September 18, 2026 release materials via Seoul Economic Daily, pending independent verification. Base for deltas is Qwen 3.5-397B unless noted.

Lineage matters: Hunmin 32B last year plus VLM 235B earlier this year precede this 397B, with Korean-language performance preserved throughout. Only-Korean-lab on the 48-model board is positioning, not capability proof, but it signals where open computer-use work now ships from. For the spend frame on serving these weights, see price per task vs price per token at 58x spread.

Architectural Impact: Grounding Plus Action Loop

Two components compose. The grounder maps screenshots plus instructions to coordinates with confidence. The action loop plans, grounds, acts, screenshots again, and verifies state changed as expected. Failures split cleanly: wrong coordinates mean grounding work, right coordinates with wrong sequence mean planning work, verified-state mismatches mean environment drift. Our dashboards now log all three separately instead of one success flag.

Eight-GPU post-training resets build-vs-buy math. Low-rank transfer of computer-operation capability plus FP8 supervised and reinforcement learning says targeted post-trains can add whole skills to open bases at lab scale. Enterprises can replicate the pattern on internal apps: capture screens plus action traces, post-train grounding adapters, evaluate on private OSWorld-style suites. Data moats beat compute moats for in-house tools. For the precision rules that keep long sessions exact, see FP8 vs BF16 vs INT4 tool-call data.

Step 1: Deploy the Open Weights

Pull the base plus FP8 variant per Xenon's release. Serve with vLLM on B200 or H100 class hardware with FP8 KV paths where validated, or BF16 KV for long tool sessions per the bakeoff above. Pin driver, vLLM commit, and dtype flags. Expose a screenshot-in plus action-out endpoint behind the same auth and metering as other agent tools, with idempotent metering guards so retried clicks bill once.

vllm serve xenon/Hunmin-VLM-397B --dtype auto --enforce-eager
vllm serve xenon/Hunmin-VLM-397B-FP8 --kv-cache-dtype fp8 --enforce-eager
python grounding_gate.py --screens eval/ours-200 --model endpoint --threshold strict

Gate first on our 200 screenshots with click-point tolerance in pixels, then on OSWorld-style task slices, then on our apps end to end. FP8 variant first for cost; fall back to full weights where grounding deltas exceed tolerance. Our threshold is zero mis-clicks on destructive actions before any production traffic.

Step 2: Migrate Tasks With State Verification

Move one workflow at a time: read-only navigation first, form fills second, destructive actions last. Each step asserts post-action screenshots against expected UI state before proceeding, with Temporal-style durable waits on human approval for irreversible clicks mirroring LangGraph on Temporal zero-crash patterns. Log coordinates, confidence, action, and verification result per step for replay.

File computer_use_policy.yaml pins allowed apps, destructive-action lists, approval scopes, and per-step timeouts. File rollout.py stages traffic 1 percent to 25 percent to full with automatic rollback on grounding-accuracy or task-success regression. No silent expansions. Every new app version re-runs the grounding gate because UI redesigns move every pixel the model memorized.

War Story 2: The Quant That Moved Every Click

Our first FP8 deploy of a vision grounder shifted click points by six pixels on dense dashboards. Buttons still highlighted, but clicks landed on borders and missed. Root cause was uncalibrated per-tensor scales plus a fused kernel enabled only for that config. We pinned eager execution, calibrated scales on our screenshots, and re-ran the gate. Accuracy recovered within a point of BF16. Lesson: grounding validation must run on your screens with your stack, never on vendor screenshots alone.

When NOT to Adopt This Release

Let us be direct. Open weights do not fit every shop.

• No GPU operations owner: 397B-class serving needs memory planning, dtype validation, and eval harnesses. Without an owner it becomes the incident.
• Browser-only workflows with stable APIs: direct API tools beat screenshots on cost, latency, and reliability. Use computer-use where no API exists.
• Regulated click paths without approval UX: destructive actions need human gates first. Add elicitation approvals before autonomy.
• Single-app pilots under 200 tasks a month: the eval harness costs more than the pilot. Start with hosted computer-use APIs.

Bottlenecks and Trade-offs

Screen resolution and scaling break grounding silently. DPI changes, zoom levels, and theme updates shift every coordinate. Pin display profiles in eval and production identically. Latency stacks per step: screenshot plus inference plus action plus verification runs seconds per click, so long task sequences need step budgets plus timeouts. Korean-language retention is claimed, not yet independently replicated; verify on your locale mix before promising coverage.

Ship Checklist

1. Grounding gate on 200 own-app screenshots with pixel tolerance, then task slices.
2. FP8 versus full-weight decision per destructive-action tolerance, stack pinned.
3. One workflow migrated with state verification plus approval scopes, staged rollout.
4. UI-version triggers re-gating, coordinates plus confidence logged per step.

Start with read-only navigation. Earn destructive clicks with data.

By Deepak Bagada, Founder and Editor-in-Chief at Daily AI World. I build agentic systems at SaaSNext and write from production logs, not demos. Follow @deeepakbagada and read more at https://deepakbagada.in.