Tencent’s Hunyuan team is reportedly gray-testing Hy4, a larger-parameter flagship model that has appeared for some users in the Yuanbao app labeled as an “expert-level model” and positioned above Hy3 and DeepSeek. Secondary reports say it can use tools to solve problems, while Tencent’s earnings materials reportedly confirm that Hy4 is in training and intended to improve performance and multimodal capabilities. Tencent has not officially released Hy4; the supplied snippets conflict on timing (“soon” versus later in 2026) and do not substantiate the claimed 770B-A49B open-weight preview or 1M-token context.
A Tencent flagship explicitly designed for tool use would converge with Scott’s view that useful agent capability combines a model with an execution surface, and would create a concrete candidate for his provider-side, trace-backed agent evaluations. The supplied evidence does not establish code-first orchestration, reliable tool performance, open weights, or practical local inference, so the connection remains prospective rather than a validation of those stronger positions.
ip:concept.model-plus-harness-benchmark-unitip:concept.agent-hands-and-eyesdev:concept.agentic-tool-loopdev:project.remote-execdev:concept.trace-backed-agent-comparisonradar:concept.frontier-modelsradar:concept.tool-useradar:concept.agent-evaluation
queries asked of Scott's wikis
- frontier models with native tool use
- agent harnesses for model-integrated tool calling
- open-weight frontier model strategy and sovereignty
- large MoE local inference economics
- evaluating expert-level models beyond benchmark claims
- multimodal models as agent foundations
2026-08-28T10:31:30Z
The prediction is absorbed: Tencent has publicly released Hy4 preview weights, completing the confirmation/release episode. Tool-use reliability, expert-level performance, licensing, and serving practicality remain separate validation questions rather than reasons to keep this release case open.
2026-08-28T09:29:17Z
Refreshed discussion only amplifies the established Hy4 preview package and adds no independent tool-use evaluation, performance validation, license clarification, or serving evidence. The release remains significant, but the immediate information spike has faded.
2026-08-28T08:38:00Z
grounded: converges/medium — A Tencent flagship explicitly designed for tool use would converge with Scott’s view that useful agent capability combines a model with an execution surface, an
2026-08-28T08:36:18Z
The added post is repetitive coverage of the already-established Hy4 preview release, reinforcing the 770B MoE, 49B-active, open-weight and 1M-context package without independently validating performance or integrated tool use. Cool the immediate release spike while watching for runnable evaluations, license details and serving results.
2026-08-28T08:23:43Z
evidence attached: reddit.post.1w0k3et — The preview reports open weights and a 770B MoE release, materially advancing the already-running Hy4 confirmation episode.
2026-08-28T07:27:40Z
Refreshed comments add architecture, context-window, benchmark, and known-issue claims while pointing to Tencent’s official WeChat post, but they do not materially change the already-alerted public preview release. Independent validation of performance, tool use, licensing, and serving practicality remains the next meaningful step.
2026-08-28T06:34:31Z
grounded: converges/medium — Tencent’s reported positioning of integrated tool use as a flagship-model capability converges with Scott’s view that tools are a core ingredient of effective a
2026-08-28T06:32:48Z
Hy4 has moved from a screenshot-derived gray-test rumor to a first-party public preview-weight release, making independent inspection and harness evaluation possible now. The release establishes the model event but not Tencent’s expert-level comparisons, integrated tool use, licensing practicality, or feasible serving profile.
2026-08-28T06:23:08Z
evidence attached: reddit.post.1w0igxk — The Hugging Face weight release is first-party artifact evidence that Hy4 has advanced beyond reported gray testing.
2026-08-27T00:29:47Z
The staleness trigger adds no evidence: Tencent’s later-2026 Hy4 release commitment remains established, but gray testing, integrated tool use, and expert-level capability are still unconfirmed. Keep the case open on a slow cadence for first-party access or independent validation.
2026-08-24T23:29:29Z
The staleness check adds no evidence beyond Tencent’s established later-2026 release commitment. Gray testing, integrated tool use, and expert-level performance remain unconfirmed, so the case should stay open but on a slower cadence.
2026-08-22T22:32:37Z
No new evidence has arrived: Hy4 remains a first-party later-2026 training and release commitment, while gray testing, integrated tool use, and expert-level performance remain unconfirmed screenshot-derived claims.
2026-08-20T21:32:59Z
The refreshed discussion adds only speculative claims about Tencent’s WorkBuddy-derived agentic data and no independent evidence that Hy4 is in gray testing, supports integrated tools, or has expert-level performance. The case remains a first-party 2026 release commitment surrounded by unvalidated pre-release interpretation.
2026-08-20T13:28:23Z
The refreshed comments remain speculative reactions to the same screenshots and add no independent confirmation of gray testing, tool use, access, or performance. The case still rests on Tencent’s narrower commitment to release Hy4 later in 2026.
2026-08-20T12:45:10Z
The new activity is engagement-only and adds no independent support for gray testing, expert-level performance, or integrated tool use. Tencent’s official commitment to release Hy4 later in 2026 remains the sole established fact, so the speculative pre-release episode has cooled.
2026-08-20T12:30:20Z
grounded: known/low — This is currently another unvalidated frontier-model/tool-use release signal of the kind already tracked in radar:deepseek-v4-flash-validation, rather than evid
2026-08-20T12:27:04Z
origin walked (codex/luna, conf 0.97): anchor reddit.post.1vth4lo -> echo.other.c1b069299b by Tencent
2026-08-20T12:25:34Z
case created — Screenshots and a linked X report indicate a bounded pre-release testing episode, but Tencent has not yet confirmed it.