On 12 November, Woshipm's "Shuzhi Benliu" column published a Hy4 hands-on test set in Tongfu Inn, the inn from the TV drama My Own Swordsman: three tasks for Hy4 inside WorkBuddy — a pelican riding a bicycle video, a find-the-difference annotation of the inn's main hall, and a Blender model of the inn built from a photo. The verdict sits in the deliverables, not in a benchmark table.
The classic model test — a pelican riding a bicycle — this time riding into Tongfu Inn: ride in from a small alley, stop clearly at the inn's entrance, delivered as a 3D-animation-style video. The first version arrived after 13 minutes: the pelican was on the bicycle and the old-town street looked the part, but the layout around the inn had changed completely — the alley that should have been a dead end now ran through both ways, and the inn's back door simply never appeared. The atmosphere of cycling through an old town was strong; the inn's exact position and the direction of its gate were a mess. A revised second version did better: sunset on blue-brick walls, the pelican riding down the alley, the inn's sign sliding in from the left, the camera settling on an old wooden door at the end of the alley — the sequence was there, but the inn itself was still not faithfully reproduced.
Testing the new visual understanding: a busy picture of the inn's main hall, full of small objects scattered across table corners, shelves and stairs, with lookalike cups, bowls and lidded containers and gold decorations mixed in with copper coins. The brief was to find each small target, count it, state its position, mark uncertainty where things were blurred or occluded, tabulate everything, and number the boxes on the original image. Hy4 returned an annotated image with 67 numbered boxes and a table of 15 inspection categories: a die on the front table edge boxed on its own, a keychain on the stair post and a key at the front of the main table annotated separately, cups, bowls and lidded containers put in different categories, copper coins distinguished from gold rings and ornaments. Checking the numbers against the image found a mix-up: box 52 circles a copper bell and box 53 circles a green vine, but the table lists "53, 56" as two copper bells and puts 52 under hanging ornaments. The author's warning is that 67 boxes do not mean 67 correctly identified objects — some boxes circle a group of shelf items, and the same object can appear in both the material category and the occlusion notes. Recognition results get written into tables, rolled into reports and handed to the next step — and the mix-up happened at exactly that step.
Given a photo of the inn's interior, Hy4 was asked to model it in Blender: judge shapes, proportions and front-back relationships from the camera's viewpoint, then place everything in 3D space. The task took 30 minutes; with no Blender installed locally, Hy4 fetched the 4.2.23 LTS portable build from an Alibaba Cloud mirror itself. The render had the feel of the inn: dark wooden beams, a second-floor corridor, lattice windows glowing warm, blue drapes along the wall, red lanterns by the door, with wood, walls and stone floors given different materials. The flaws were real too: a staircase with a landing became a single straight flight, changing its orientation and the proportions of the surrounding space; the lattice windows repeated one pattern; the lanterns and furniture were simplified. Asked to move only the counter to the left and keep everything else in place, then unfold the same way: the counter with its abacus and wares moved to the other side, the old spot left empty, stairs, tables, chairs, the gatehouse and the roof unmoved, the unfolding order and camera rotation preserved.
From the start of testing to the end, 82 credit records totalled 3,429.15 credits; 48 of them cost nothing (44 hy4-free and 4 wb-lite). Video generation accounted for 26 charges and 3,382.07 credits — 98.6% of the total (model tag minimax-video-h3); image editing took 4 charges for 23.08 credits, and the remaining 24 credits went to other model calls, including an automated news task. The author's conclusion: in these tasks, the expensive part was not the model, it was making things move — intelligence itself has got cheap, rendering and generation have not. On pricing itself: Hy4's full release keeps the preview's list prices (¥6 in, ¥18 out, ¥0.3 cached), Tencent Cloud's 50% discount runs to 31 December, and WorkBuddy and CodeBuddy offered some free quota from 4 to 15 November.
The official and independent figures behind this test are already mapped in our benchmarks article: WorkBuddy office evaluation averaged 83.18 with a 96.15% pass rate, No. 1 in China in 9 of 12 scenarios; 3.04 on 579 real finance tasks, best among open models; AA index 35, No. 2 in China; FrontierCode 2.0 at 31%, No. 1 among Chinese models; and on AA's knowledge-reliability benchmark, an 89% rate of answering firmly when it does not know. What the Tongfu Inn tasks add is what no benchmark table gives: spatial consistency and the reliability of tabulated output.
Reference: Woshipm — How Did Tencent Hunyuan Hy4 Do Working at Tongfu Inn? (Chinese) (rewritten and organized by ModelBrief)