Hy4 Official Launch: Vision & Computer Use Added, Price Frozen — Should You Migrate?

Core Upgrade: Sensory Capabilities, Not Architecture

Hy4 official retains the 770B MoE architecture from the August preview; changes are expanded pre-training and post-training data and scale. Two net-new capabilities appear:

Vision understanding : fine-grained image recognition, complex chart analysis, math reasoning with images — e.g., feed a financial-report screenshot or a fridge photo directly.

Computer Use : combines command-line and GUI interaction, with specialized training on Blender, CAD, and cross-application long-horizon tasks.

In short, preview was a "read-write model"; official is a "see-and-act model." Upgrade only matters if your workloads involve image inputs or autonomous software operation.

Official vs Preview: Key Deltas

⚠️ Open-source status regressed. Preview weights were released August 28 on HuggingFace/ModelScope; official only promises "weights and system card coming soon." Private deployment today still runs preview, not the stronger official version.

⚠️ Benchmark numbers need discounting. Hallucination rates differ across sources: 12.95%→7.2% (Beijing News), 15.76%→5.45% (Sina Finance); only "guess-induced error rate" 23.1%→5.6% is consistent. Most cited evaluations (financial crowd-test with 213 experts/579 tasks, WorkBuddy office benchmark, TenureTrail research benchmark) are Tencent-organized, not independent. Tencent admits Hy4 still trails GPT-6.1 Sol and Claude Opus 5.5 on ProgramBench, FrontierCode 2.0, Agents' Last Exam v1.1, though it beats DeepSeek-V4.1-Flash-1017 on some.

As the author notes: "Benchmarks are the vendor's map, not your road conditions" — measure your own task failure rates before migrating.

Cost Analysis: Cheap Tokens ≠ Cheap Bills

Pricing unchanged: input 6¥/M, output 18¥/M, cache hit 0.3¥/M. For a mid-size team (10M tokens/day, 80/20 input/output split, 30 days = 300M tokens):

Input: 240M × 6¥ = 1,440¥/month

Output: 60M × 18¥ = 1,080¥/month

Total ≈ 2,520¥/month; with TokenHub 50% off until Dec 31 → ~1,260¥/month.

Reference: GPT-6.1 Terra at $1/$5 per M tokens (~7.2/36¥ at 7.2 FX). Hy4 is cheaper per token, but two hidden costs dominate:

Token consumption spikes with vision/computer-use: a high-res image, a GUI trajectory, or a long agent loop can multiply context 10-50× over pure text. Per-task cost must be measured empirically.

Migration & acceptance testing : new modalities change output behavior; a model-id swap takes minutes, regression testing takes weeks.

Free-trial window: WorkBuddy/CodeBuddy offer limited free quota Nov 4-15; Hy4 flash (lower cost) promised "soon." High-volume cost-sensitive teams should wait for flash and benchmark per-task economics first.

Adoption Decision Matrix

Cloud teams : adopt now via canary + own evaluation set; capture 50% discount and free quota.

Private-deployment teams : wait for official weights and system card; do not substitute preview.

Local/on-device teams : Hy4 770B still requires 8-GPU tensor parallel; no Mac (even 512GB M5 Ultra) can run it. Stick with Qwen3.8-27B or await Qwen4-27B.

Cost-sensitive high-volume teams : measure per-task cost first, then wait for Hy4 flash.

Three Caveats Before Migrating

Evaluation scope : mostly vendor-run, dual hallucination numbers, no independent replication. Decide on your own task suite, not the vendor's 83.18 score.

Admitted gap : Tencent acknowledges "large gap" to top closed models. If your workload already runs well on GPT-6.1 Sol/Claude Opus 5.5, don't migrate for "domestic/open-source" labels alone.

New capabilities = new risk surface : computer use introduces permission boundaries, mis-operation, runaway long tasks; vision adds prompt-injection vectors. The author's vertical-agent prerequisites — boundaries, acceptance loops, tooling — must be in place before letting a model click a mouse.

Data sources: Tencent Hunyuan Nov 4 2026 release, Beijing News, Tencent News, The Paper, Sina Finance, Jishi Tech; cost estimates at list prices; benchmarks largely vendor-organized; hallucination rates from two differing methodologies.