Tencent open-sources Hy4 preview (770B MoE) — free for two weeks in CodeBuddy, but the benchmarks are internal

2 sources

Tencent released and open-sourced Hy4 preview on 2026-08-28 — a mixture-of-experts model with 770B total parameters (49B active) and a context window exceeding 1M tokens, aimed at software engineering, office productivity, and scientific research. It is reachable in-product through WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access via Tencent Cloud TokenHub and OpenRouter. Hy4 preview is free to use on WorkBuddy and CodeBuddy for two weeks from launch, and free access to the previous Hy3 is extended until September 30. Tencent lists preview API pricing at USD 0.834 per million input tokens and USD 2.501 per million output tokens (USD 0.042 per million cache-hit tokens). Tencent also reports an internal blind evaluation — 163 experts scoring 203 real engineering tasks — where Hy4 preview averaged 2.99/4.00 versus GLM-5.3 at 2.92 and Kimi K3 at 2.94.

Why now

The free two-week window on WorkBuddy and CodeBuddy opened on launch day (2026-08-28), so creators can try the model at zero cost this week — and Hy3 free access ends September 30, giving the piece a natural before/after deadline. Preview pricing is published, so the cost math is quotable on camera. The headline comparison, though, is Tencent's own blind test — which makes the honest-read angle as timely as the release itself.

Why it is worth publishing

Free access plus published pricing makes the demo executable today; the internal-eval caveat gives the video a trust angle that most launch-day coverage will skip. Reuters' same-day report supplies the independent media anchor, so this card carries the Independently verified badge.

Evidence basis

Tencent's official announcement (2026-08-28, opened and read twice) + Reuters same-day coverage (2026-08-28).

“Tencent just open-sourced a 770-billion-parameter model and made it free for two weeks in CodeBuddy — here's what it's actually good at, and which benchmark numbers to take with a grain of salt.”

Angle

Lead with the access window, not the parameter count: a 770B open-weights model is a vendor flex, but 'free in CodeBuddy for two weeks' is something a creator can act on today. Pair it with the honest read of Tencent's internal blind test.

Format

Short talking-head video

Demo idea

Run one real office task end-to-end on Hy4 preview inside WorkBuddy or CodeBuddy — a doc-to-spreadsheet job or a small front-end build — then compute the same job's API cost at USD 0.834/2.501 per million tokens on screen. Close with the caveat: the 2.99-vs-2.92 comparison comes from Tencent's own blind test.

Platform notes

Attribute specs and pricing to Tencent's 2026-08-28 announcement and link it. Say 'Tencent reports' before any head-to-head score — the comparison with GLM-5.3 and Kimi K3 comes from Tencent's internal blind evaluation, not an independent benchmark.

Usable claims

Tencent released and open-sourced Hy4 preview on 2026-08-28: a mixture-of-experts model with 770B total parameters, 49B active parameters, and a context window exceeding 1M tokens, aimed at software engineering, office productivity, and scientific research. Tencent states the model is available as an open-source release and reachable in-product through WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access through Tencent Cloud TokenHub and OpenRouter; Reuters' report adds that the release is via Hugging Face.
Tencent's announcement lists preview API pricing for Hy4 preview at USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million cache-hit tokens. Hy4 preview is free to use on WorkBuddy and CodeBuddy for two weeks from launch, and free access to the previous Hy3 has been extended until September 30.
Tencent reports an internal blind evaluation in which 163 experts scored 203 real engineering tasks: Hy4 preview averaged 2.99 out of 4.00, ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. The evaluation was run by Tencent itself; Reuters reports the same comparison as Tencent's own claim, not as an independent benchmark.
Tencent states Hy4 preview participated in automated optimization of its own training methods, data strategies, evaluation frameworks, and low-level operators — proposing approaches, running experiments, and iterating — which Tencent describes as an early-stage recursive self-improvement loop. Tencent also credits autonomously optimized operator fusion and communication for a 31.8% end-to-end inference throughput gain over baseline. Both statements are vendor-reported.

Evidence pipeline

From the news

Breakdown

Tencent's 2026-08-28 announcement open-sources Hy4 preview (770B total / 49B active MoE, >1M-token context) with a published preview-price list (USD 0.834 input / 2.501 output / 0.042 cache-hit per million tokens) and a two-week free window on WorkBuddy and CodeBuddy. This breakdown separates the checkable facts (release date, parameter counts, pricing, access paths) from the vendor-reported layer (the 163-expert blind test scored 2.99/4.00 against GLM-5.3 and Kimi K3, and the recursive self-improvement narrative) — and shows how to attribute each on camera. Reuters' same-day report is the independent anchor on the release event, including Tencent's own preview-maturity caveat.

Sources

Risks

Label every demo as Hy4 preview with the date on screen. Re-run the same task after the official release and note what changed. Don't bake preview-only behavior into an evergreen tutorial.
Attribute on screen: 'Tencent says its internal blind test scored...'. Keep the numbers next to the attribution, and pair them with an independent benchmark if one exists by the time you publish.

Demo ideas

Zero-cost window walkthrough: one real doc-to-spreadsheet task in CodeBuddy on Hy4 preview, with the launch-window and Hy3-September-30 dates on screen
Cost math card: same task priced at USD 0.834 input / 2.501 output per million tokens, labeled preview pricing