DK Dark Knight 2026-08-29 5 min read

Tencent Dropped 770B of Open Weights. Almost Nobody Can Run Them.

Tencent dropped Hy4 preview on August 28: 770B total parameters, 49B active, a 1M-token context window, Apache 2.0. On paper it's the strongest open-weight release of the year, and herding it onto Hugging Face took 131 shards and 1.56TB of BF16 weights. The community responded in hours — GGUF and MLX 4-bit ports are already live, and the official repo passed 250 stars in its first two days.

I spent the morning inside the release — the model card, the config.json, the safetensors headers. The interesting number isn't 770B. It's the machine you need just to get the weights into VRAM. Spoiler: it is not a machine you own.

Open-source AI's real price moved from the license to the serving contract the day a 770B model hit Apache 2.0.

What actually shipped

I opened the box

The release is 1.56TB and I don't have that. But you don't need it to verify a surprising amount — the config and the safetensors index are plain JSON, and a shard's header is a JSON blob sitting at byte 8 of the file.

# architecture & shape from config.json
$ python3 -c "import json; d=json.load(open('hy4_config.json')); \
  print(d['architectures'], d['num_hidden_layers'], 'layers,', d['n_routed_experts'], 'routed experts')"
['HYV4ForCausalLM'] 78 layers, 256 routed experts

# 2006 named tensors spread across 131 shards
$ python3 -c "import json; print(len(json.load(open('hy4_index.json'))['weight_map']))"
2006

# peek at shard 1's header via a 1MB range request
$ curl -r 0-1048575 tencent/Hy4-preview/resolve/main/model-00001-of-00131.safetensors
> 2 tensors, dtype BF16
> model.layers.1.mlp.experts.down_proj   shape [256, 6144, 2048]

That single tensor is 3.2B parameters. 77 MoE layers × 256 experts × 3 matrices ≈ 743B — add attention, embeddings, and the MTP head, and the 770B claim checks out to the digit. The weights are what they say they are.

Here's the catch: almost nothing can load it yet

The card's quickstart runs into transformers_version: 5.16.2. PyPI's latest is 5.16.1 (Aug 26). Transformers main has hunyuan_v1_dense and hunyuan_vl in the auto-mapping — but no HYV4ForCausalLM, no hy_v4 anywhere. A stock from_pretrained fails today, full stop.

vLLM is the same story with better discipline. The official recipe demands vLLM 0.29.0 and targets "16×B200, 8×B300 with MTP." The latest release tag is v0.28.1rc0, and v0.28.0's model list — published three days before the weights — has no Hy4. The only working path right now is a prebuilt image:

docker run --gpus all -p 8000:8000 --ipc=host \
  -e VLLM_ENABLE_HPC_OPS=1 \
  vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 \
    --tensor-parallel-size 8 \
    --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' \
    --attention-backend FLASHMLA_SPARSE \
    --tool-call-parser hy_v4

SGLang ships the equivalent prebuilt image with NEXTN speculative decoding. Look at what's bundled: not just weights, but custom kernels (FLASHMLA_SPARSE), an HPC-ops environment flag, and tool-call/reasoning parsers for hy_v4. That is the real release. The weights are the entry ticket; the serving stack is the product. Tencent shipped the whole bundle before the ecosystem had merged anything — the merge train (vLLM 0.29.0) is still catching up.

What "open source" means at 770B

Run the memory math as an individual:

No single machine an individual can buy serves this thing at useful quality. The cheapest honest deployment is a five-figure monthly rental bill, and the officially supported configuration costs more than most startups' runway. Apache 2.0 is a license for people who already own the hardware. For everyone else, the actual product is the API — free for two weeks on WorkBuddy/CodeBuddy, then via Tencent Cloud TokenHub and OpenRouter (the API rollout is covered in the earlier post today).

graph TD
  IN[Tokens] --> EMB[Embed · vocab 120,832]
  EMB --> L0[Layer 0 · dense FFN]
  L0 --> MOE[Layers 1–77 · MoE]
  MOE -->|top-8 of 256 + 1 shared| EXP[Expert FFN · 2,048]
  MOE --> ATT[Gated DSA + MLA
q-lora 2,048 · kv-lora 512 · IndexCache]
  EXP --> IHC[iHC residuals · 4 streams]
  ATT --> IHC
  IHC --> OUT[LM head · 770B total / 49B active]
  MOE -.->|MTP head · 10B| SPEC[Speculative decoding · 3 tokens]

Bottom line

Hy4 preview is real, the weights are honest, and Tencent did something genuinely useful: weights, kernels, images, and recipes released as one coherent bundle under a permissive license. But read the Apache 2.0 as a tax on the already-equipped, not an invitation. The frontier of "open" is no longer defined by who can download — it's defined by who can serve. That's a quieter, more expensive gate, and it's worth being honest about.

Related reading