Tencent's Hy4-Preview Is a 770B Open Model That Leads SWE-bench Pro
Quick verdict
Tencent put out Hy4-preview and called it "open source frontier," which is the kind of line labs use loosely. This time the numbers back it. Hy4 is a 770-billion-parameter mixture-of-experts model with 49B active and a 1M-token context window, and it did not land as a quiet checkpoint bump. It jumped 115 points over Hy3 on Code Arena's WebDev board and leads SWE-bench Pro on early runs. If the coding scores hold, the open-weight frontier just moved again, and it moved on the axis Admix readers care about most: shipping code.
What actually shipped
The Tencent Hunyuan account released Hy4-preview with 770B total parameters, 49B active per token, and a 1M-token context window, framing it explicitly as an open frontier model rather than a research artifact. The size alone puts it near the top of the open MoE field, next to GLM-5.3 and the newest Qwen releases.
The external signal is stronger than the spec sheet, though. The Arena account placed Hy4 around #5 on Code Arena: WebDev through AutoEval, a 115-point jump over Hy3. That is not an incremental refresh. On the agent side, Cline reported that Hy4 leads SWE-bench Pro, the harder, contamination-resistant version of the benchmark that tracks real pull-request fixes. Put those together and Hy4 is not competing for a spot on the open leaderboard, it is competing with closed coding models.
The Codex-orchestration claim
The detail that got engineers talking was not a benchmark. Per a widely shared summary, Tencent claims Hy4 can coordinate multiple Codex sessions in parallel for research workflows. That reframes the model from "thing that writes code" to "thing that runs other coding agents," which is exactly where the tooling has been heading. If a single open model can act as the orchestrator over a fleet of agents, the cost math for agentic coding shifts, because the expensive planning layer no longer has to be a closed API call.
Treat the orchestration claim as promising rather than proven. Coordinating parallel agents is far harder than topping a static benchmark, and Tencent's own framing is a capability pitch, not an independent result. Still, the direction lines up with the broader move toward the harness becoming part of the model system rather than a wrapper around it.
Why the serving design matters
Open weights nobody can run cheaply are a press release. Hy4's architecture reads like it was built to avoid that. The vLLM project flagged an unusually deliberate serving design: 256 routed experts plus 1 shared expert, with only 21 of 78 layers computing their own sparse routing index while the rest reuse it, plus an embedded 10B multi-token-prediction layer running at draft depth 3. In plain terms, Tencent shaved the routing overhead that usually makes giant MoE models slow to serve, and baked in speculative decoding rather than bolting it on later. That is the difference between a model providers can host at a sane price and one that stays a demo.
Where it sits against the open field
Hy4 lands in a crowded week. GLM-5.3 open-weighted its flagship, Qwen shipped a cheap long-context Flash variant, and the open frontier keeps compressing the gap to closed models. What separates Hy4 is the coding lead specifically. Hy3 was a strong second option next to GLM; Hy4 is making a case to be the open model you reach for when the task is a real codebase. We tracked the incumbent in our GLM-5.2 breakdown and the earlier Tencent release in the Hy3 writeup, and Hy4 is the first open model in a while that changes the ranking on coding rather than just adding another name to it.
The catch: 770B is not a laptop model
None of this makes Hy4 something you run at home. 49B active parameters per token help throughput, but you still have to hold all 770B in memory to route across 256 experts. For nearly everyone, "open" here means rent it from an inference provider, not spin it up on a workstation. The value of the open license is that no single vendor controls the weights, so several providers can serve the same model and you are not locked into one company's pricing or uptime. If you want to compare Hy4 against closed models without paying for six subscriptions, running them side by side in one app that hosts multiple AI models is the cheaper path.
Video: Hy4-preview tested
An early hands-on look at Hy4, how it stacks up on coding, and how to try it while the free window is open.
FAQ
Can I run Hy4 on my own machine?
Realistically, no. It is a 770B mixture-of-experts model, so even though only 49B parameters activate per token, you need enough memory to hold all 256 experts. Most people will use it through a hosted provider. Our roundup of open-source AI models covers what is actually runnable locally.
Is Hy4 better than GLM-5.3 for coding?
Early signals favor Hy4 on coding specifically. It leads SWE-bench Pro on Cline's runs and jumped 115 points over Hy3 on Code Arena WebDev. Both models are new, so treat this as a strong early lead rather than a settled result, and check independent tests before betting a workflow on it. See our best AI models for coding guide for the wider field.
What does the parallel-Codex claim actually mean?
Tencent says Hy4 can coordinate several Codex coding sessions at once for research workflows, acting as an orchestrator over other agents rather than just generating code itself. It is a capability claim from Tencent, not an independent benchmark, so it is worth watching rather than banking on.
Sources
- @TencentHunyuan - Hy4-preview release: 770B total, 49B active, 1M context, "open source frontier"
- @arena - Hy4 around #5 on Code Arena WebDev, a 115-point jump over Hy3
- @cline - Hy4 leads SWE-bench Pro
- @kimmonismus - Tencent's claim that Hy4 coordinates multiple Codex sessions in parallel
- @vllm_project - Hy4 serving design: 256 routed + 1 shared expert, 21/78 layers, embedded 10B MTP layer
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix