Tencent's Hy4 Preview attracted 228 points and 126 comments on Hacker News within hours, yet the number developers will feel first is 814 GB. That is roughly how much storage the FP8 model repository occupies, and Tencent's recommended vLLM command splits it across eight GPUs. The release is open under Apache 2.0 and can process a claimed 1 million tokens, but self-hosting the flagship is a data-center job rather than a weekend laptop experiment.
The scale makes this a useful test of what an open model release means in 2026. Tencent has published both BF16 and FP8 weights, deployment instructions, fine-tuning material and an OpenAI-compatible serving path. Developers can inspect and modify the model without waiting for access from a closed API vendor. Most teams, however, will meet Hy4 through a hosted endpoint because downloading and serving it demands serious infrastructure.
A 770-billion-parameter mixture of experts
Hy4 Preview's model card lists 770 billion parameters in its backbone, with 49 billion activated for each token. It has 78 layers. The first uses a dense feed-forward network, while the remaining 77 use a mixture-of-experts design with 256 routed experts and one shared expert per layer. Each token selects eight routed experts alongside the shared one.
That arrangement separates the model's total capacity from the amount of computation used for one token. It does not make the other weights disappear. The BF16 repository uses about 1.56 TB of storage across 131 safetensor shards, according to the Hugging Face repository metadata. The FP8 release cuts that figure to about 814 GB, still far beyond a typical workstation's GPU memory.
Tencent also includes a 10-billion-parameter multi-token prediction layer, of which 700 million parameters are active. It is intended for speculative decoding, where the system proposes several future tokens and verifies them in batches. The supplied vLLM recipe enables three speculative tokens and names a dedicated sparse-attention backend. Those details matter because the headline parameter count says little about whether a model can produce answers at a usable rate.
Hy4's attention system draws on DeepSeek Sparse Attention and an approach called IndexCache. The model card says the indexer selects 2,048 positions and reuses sparse indices across layers. Its configuration sets the maximum position count to 1,048,576 tokens. Tencent describes that as a 1-million-token context window, though the release does not provide application-level tests showing accuracy across an entire window. Capacity and reliable recall are separate measurements.
The official local route starts with eight GPUs
Tencent provides prebuilt images for vLLM and SGLang. The published vLLM command loads the FP8 checkpoint and sets --tensor-parallel-size 8, while the SGLang example also uses a tensor-parallel size of eight. Neither recipe names a minimum GPU model or a tested memory floor, so it would be unsafe to turn that command into a universal hardware recommendation. It does establish the intended deployment class.
The serving layer exposes an OpenAI-compatible API. Once an operator has the model running, an existing client can point at a local base URL and request hy4-preview. Tencent recommends a temperature of 0.9 and makes high reasoning effort the default; callers can pass a no_think setting for direct responses. That API compatibility lowers migration work, even though it does nothing to reduce the cost of hosting the weights.
The Apache 2.0 license is unusually permissive for a model of this size. Tencent's license file permits use, modification and redistribution under the license's notice and attribution conditions. The repository also includes scripts for full and LoRA fine-tuning through DeepSpeed, LLaMA-Factory and ms-swift. Access to weights therefore has practical value for research groups and infrastructure companies that can supply the hardware.
For everyone else, Tencent is offering Hy4 through WorkBuddy, CodeBuddy and several of its other products. The company says API access is available through Tencent Cloud TokenHub and OpenRouter at $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached tokens. Those are Tencent's announced prices, not a measurement of total workload cost; reasoning length, cache use and tool calls can change the bill substantially.
Tencent's strongest results are still Tencent's results
The company positions Hy4 as a productivity model for software work, office documents, game development and scientific research. Its most concrete human evaluation used 163 Tencent experts and 203 engineering tasks. Tencent reports an average score of 2.99 out of 4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Against GLM-5.3, it reports 46.8 percent wins, 12.8 percent ties and 40.4 percent losses; against Kimi K3, 51.2 percent wins, 7.9 percent ties and 40.9 percent losses.
Those figures come from an internal blind evaluation described by the model maker. The model card does not publish the 203 prompts, the complete outputs, grader instructions or per-task scores. The small differences in the averages also leave little room for sweeping conclusions. Independent tests will need to reproduce the comparisons and examine whether the gains hold outside Tencent's product workflows.
Tencent acknowledges two early problems in the preview. The model can spend longer than necessary reasoning through complex tasks, and it can over-verify its own work. Both behaviors affect latency and token consumption, especially when high reasoning effort is the default. Calling the release a preview is meaningful here: the company is shipping known behavioral issues and asking users to report failures before the final Hy4 line arrives.
The release announcement makes a more unusual claim about development. Tencent says Hy4 helped optimize its own training methods, data strategies, evaluation systems and low-level operators. According to the company, the model proposed approaches, ran experiments and fed code, logs and feedback into later iterations. It also says the model analyzed inference bottlenecks and helped produce a 31.8 percent end-to-end throughput increase over an unspecified baseline.
That 31.8 percent result lacks enough public detail for an outside performance comparison. Tencent says the gain held across context lengths and concurrency levels, but it does not disclose the hardware, serving configuration, baseline build, traffic mix or absolute token rates in the announcement. The work may be technically useful; the published evidence only supports describing it as Tencent's measured improvement in its own system.
What the release changes for developers
Hy4 adds another very large model whose weights can be downloaded, audited and adapted under a standard software license. It also arrives with working recipes for two widely used serving engines instead of a model card that stops at architecture diagrams. That makes evaluation possible for labs, cloud providers and companies already operating multi-GPU inference clusters.
The release is less accessible to individual developers than the word "open" may suggest. An 814 GB quantized artifact takes time to transfer and store before GPU memory enters the calculation. Eight-way tensor parallelism also brings interconnect, scheduling and reliability concerns. A hosted API avoids that operational burden but gives up some of the control that makes downloadable weights attractive, including local data handling and the ability to modify the serving stack.
The next useful evidence will come from public evaluation artifacts and independent serving reports. Watch for Tencent to disclose the engineering-task set, specify the setup behind its 31.8 percent throughput claim, and publish tested hardware profiles for the FP8 build. Also watch whether later Hy4 releases reduce unnecessary reasoning without losing task accuracy. Until those details arrive, Hy4 Preview is best read as a genuinely open, unusually large model release with a clear operational price attached.