Hi-4 Preview Hands-On: Coding, 3D Games, and Document Audits
Hands-on tests of Tencent's Hi-4 Preview model tackling a platformer, a 3D racing game, and a 24-claim expense audit in one shot.
What is Hi-4 Preview and why does the hands-on testing matter?
Hi-4 Preview is Tencent’s newly released open-weight model, a 770 billion parameter mixture-of-experts system with 49 billion active parameters and a context window over 1 million tokens. Benchmarks tell you what a model scores on a test set. Hands-on testing tells you whether it can actually finish a real task without falling apart halfway through. A recent round of demos put Hi-4 Preview through three tasks that stress different muscles: one-shot game coding in 2D and 3D, and a long, multi-document expense audit. All three were run inside WorkBuddy, a desktop agent app, using single prompts with no follow-up corrections.
TL;DR
- Hi-4 Preview is a 770B-parameter mixture-of-experts model with 49B active parameters, released under Apache 2.0 with weights on Hugging Face in BF16 and FP8.
- It one-shotted a Super Mario style platformer built with raw HTML5 canvas and vanilla JavaScript, including working physics, collision detection, and hand-drawn pixel art with no image assets.
- It also one-shotted a 3D neon street racer in Three.js with drifting, nitrous effects, AI opponents, live standings, and a lap timer, a task where a comparison model dropped the AI opponents and nitrous entirely.
- On a 24-claim expense audit spanning policy documents, employee records, invoices, and emails, it classified every claim, applied the correct policy version by date, and caught duplicate invoice numbers and an over-used allowance.
- Pricing undercuts most open-weight flagships at $0.834 per million input tokens and $2.001 per million output tokens, with cache hits at $0.042 per million.
- The model needs serious hardware to self-host (BF16 weights around 1.8TB, FP8 around 900GB), so most people will use it through an API or an app rather than running it locally.
- Tencent claims Hi-4 Preview helped optimize its own training pipeline, improving throughput by roughly 31.8% against baseline.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How does Hi-4 Preview perform on coding tasks?
The first test asked the model to build a Mario-style platformer using only the native HTML5 canvas API and vanilla JavaScript, with no game engine and no image assets. The requirements were specific: a requestAnimationFrame game loop, full collision detection for terrain and item blocks, mushroom and coin drops, enemy stomping, and physics covering gravity, inertia, run acceleration, and variable jump height.
Hi-4 Preview responded by breaking the task into stages first (game loop, physics constants, tile map, collision system, entities, rendering) before writing code. It then opened the finished file in a browser on its own to verify it ran. In testing, the result held up: acceleration and inertia behaved correctly, jump height varied with key hold duration, item blocks produced a mushroom that slid along the ground, and enemy stomping and brick collision from both above and below worked as specified. All artwork was drawn directly with canvas calls rather than sprite images. The task was completed in a single prompt with no corrections needed, which is notable because collision detection is a common failure point even for larger closed models attempting similar builds.
Can it handle 3D game development?
The second test pushed further into 3D: a Three.js neon street racing game with a third-person chase camera, wet reflective streets, glowing neon signage, motion blur, a nitrous boost with a widening field of view and colored exhaust flames, AI-controlled opponents, live race standings, and a three-lap timer, all specified in one prompt.
The model again planned before building, laying out the scene, lighting, road geometry, car physics, opponent AI, and HUD as separate stages. The output included working wet-road reflections, glowing neon materials, a chase camera that tracked the car properly, functional drifting, a nitrous bar that filled with use, and a boost effect that widened the camera’s field of view while triggering blue and purple flame effects. AI opponents raced independently and the on-screen standings updated as positions changed during the run.
For comparison, the same prompt was run against a different model inside the same app. That model produced basic driving mechanics but skipped the AI opponents entirely and left the nitrous effect non-functional. The gap suggests Hi-4 Preview handles prompts with many simultaneous constraints more reliably, at least in this specific test, rather than dropping requirements as the prompt gets longer.
How well does it handle long, multi-source document work?
The third test moved away from code entirely. The setup: a folder containing 24 reimbursement claims, two versions of a policy document, employee records, budget files, allowance usage logs, invoice history, and supporting emails. The prompt asked the model to act as an expense audit specialist, checking every claim against all supporting materials, classifying each as approved in full, approved with deductions, or returned for more documentation, then producing a summary with evidence-based findings, fraud flags, total deductions, and a list of missing documents.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
This is the kind of task that typically takes a person the better part of a day. The model read through the full folder, correctly identified that two policy versions applied depending on claim date, and worked through each claim while cross-referencing invoices, emails, and the budget file. The final output was a written summary document plus a spreadsheet with per-claim classifications, amounts, deductions, the specific policy clause applied, and the evidence used to justify each decision.
Three claims were flagged as likely duplicates because the same invoice number appeared twice with different dates. One claim was flagged because the employee’s allowance had already been exhausted according to the usage log. Two more were caught for being submitted under an outdated policy version after the new one took effect. The total deduction figure at the bottom of the spreadsheet checked out against manual verification, and the model stayed consistent across all 24 claims without inventing evidence, which is typically the hardest part of long-horizon document tasks.
Is Hi-4 Preview worth using over other open-weight models?
For agentic tasks like coding, game development, and document-heavy audits, the testing suggests Hi-4 Preview is competitive with or ahead of other current open-weight models such as Kimi K3 and GLM 5.3, based on Tencent’s internal blind evaluation involving 163 experts across 203 real engineering tasks. Take that internal number with some skepticism since Tencent ran it, but it lines up with independent benchmark results: 92.3 on GPQA Diamond, 85.4 on Terminal-Bench, and 82.9 on SWE-Bench Multilingual, putting it at or near the top of the open-weight field.
It still trails closed frontier models like Opus 5 and GPT 5.6 on the hardest tasks, which is expected for an open model. But the pricing is a real advantage: at $0.834 per million input tokens and $2.001 per million output tokens, it undercuts Qwen 3.8 Max and Kimi K3, though DeepSeek V4 Flash remains cheaper for anyone optimizing purely for cost. Being released under Apache 2.0 also means third-party providers can host it freely, which matters more for availability and pricing competition than raw benchmark position.
Can you run Hi-4 Preview locally?
Not practically. The BF16 weights come in around 1.8TB, and even the quantized FP8 build needs roughly 900GB of memory, meaning a minimum of several high-end GPUs (Tencent’s own reference points to something like 16 B200-class cards) just to load it. For almost everyone, that rules out local inference. The realistic paths are Tencent Cloud’s Token Hub, OpenRouter for plugging into existing coding agents like Claude Code or other CLI tools, or a desktop app like WorkBuddy, which offered the model free for two weeks after launch as a way to let people test it without committing to API costs.
Frequently Asked Questions
What is Hi-4 Preview built for?
It’s designed for agentic work rather than general chat: coding agents, multi-step planning, long-horizon task execution, tool calling, and office productivity tasks like generating spreadsheets and slide decks.
How big is Hi-4 Preview and what license does it use?
It’s a 770 billion parameter mixture-of-experts model with 49 billion active parameters per token, using 256 routed experts plus a shared expert. It’s released under the Apache 2.0 license with weights available in BF16 and FP8 on Hugging Face.
Can Hi-4 Preview really one-shot a working game?
One coffee. One working app.
You bring the idea. Remy manages the project.
In hands-on testing, yes for both a 2D canvas-based platformer and a 3D Three.js racing game, including physics, collision detection, and visual effects, all from single prompts without follow-up fixes.
How does Hi-4 Preview compare on price to other open models?
It costs $0.834 per million input tokens and $2.001 per million output tokens, with cache hits at $0.042 per million, undercutting Qwen 3.8 Max and Kimi K3, though DeepSeek V4 Flash is still cheaper.
Do I need special hardware to run Hi-4 Preview myself?
Yes, and most people won’t have it. The BF16 weights are about 1.8TB and the FP8 build needs roughly 900GB of memory, requiring multiple high-end GPUs. Most users will access it through an API, OpenRouter, or an app rather than self-hosting.
