Tencent Hy4 Preview Review: Tencent's 770B Open Reasoning Flagship
Tencent released and open-sourced Hy4 preview on August 28, 2026 — a 770B-total / 49B-active Mixture-of-Experts reasoning model with a 1M+ token context, Apache-2.0 weights, and a launch blind test in which 163 engineering experts rated it ahead of GLM-5.3 (2.99 vs 2.92 out of 4). It is Tencent's answer to the August open-model wave — and, true to its name, a preview: the company positions it as the first of more Hy4 models to come. This review separates what shipped, what the evidence actually shows, and what "preview" means for adoption.
Preview vs. full release — read this first
Last updated: August 30, 2026. Tencent explicitly labels this release preview: a deliberately early cut of the Hy4 generation, open-sourced for community testing, with more Hy4 models promised to follow. Public numeric benchmark scores were not released at launch (the GitHub repo's benchmark results are chart images, not tables). The headline evaluation is a blind expert test described in Tencent's own announcement and repo. Treat everything below accordingly.
Quick Verdict
Hy4 preview is the biggest open model of the August 2026 wave and the first Tencent flagship released open since the company committed to the open-source path: 770B parameters, 49B active, a real 1M+ context, Apache-2.0, and a credible (if vendor-run) expert evaluation that edges GLM-5.3 on engineering tasks. It also ships with the least numeric benchmark evidence of the week's three open releases — no public coding/math scores, a known over-reasoning tendency, and a "preview" label that Tencent itself uses to manage expectations. For teams that want the most capable Apache-2.0 reasoning model available today and can tolerate early-release roughness, it is a serious contender; for teams that need benchmark tables, the wait is short — the full Hy4 line is explicitly coming.
What Is Hy4 Preview?
Hy4 preview is a new-generation Mixture-of-Experts flagship model developed by the Tencent Hy Team, announced and open-sourced August 28, 2026 (Chinese media reported Tencent confirming plans for the release from around August 12). It follows Hy3 as the "expert-grade" upgrade with larger scale, stronger reinforcement learning, and multimodal gains per Tencent's positioning. The model is built for "productivity" scenarios: AI coding, office and financial analysis, game development, and scientific research. The "preview" designation is deliberate — Tencent describes it as the first step of the Hy4 generation, with subsequent Hy4 models to follow soon.
Model Specs & Architecture
| Spec | Hy4 preview |
|---|---|
| Release date | August 28, 2026 (announced/open-sourced) |
| Total parameters | 770B (backbone; +10B in native MTP layer) |
| Active parameters / token | 49B (backbone; +0.7B MTP) |
| Architecture | MoE, 78 layers: 1 dense FFN + 77 layers of 256 routed experts + 1 shared expert, top-8 activated |
| Attention | Gated DSA with IndexCache; query compression 2048, KV compression 512, 32 indexer heads |
| Context window | 1M+ tokens (official newsroom: "over 1M") |
| Hidden size | 6144; 64 attention heads; vocab 120,832 |
| Speculative decoding | Native MTP layer (10B total / 0.7B active) |
| License | Apache-2.0 |
| API price (in / out / cached, per 1M) | ¥6 / ¥18 / from ¥0.3 (≈$0.85 / $2.54 / $0.04) |
Architecture details from the official repository: the backbone's 77 expert layers each route through 256 routed experts plus one shared expert, activating the top-8 per token; a native MTP (multi-token prediction) layer adds 10B total / 0.7B active parameters for speculative decoding; and the Gated DSA attention with IndexCache compresses queries (2048) and keys/values (512) to keep the 1M context tractable. Two optimizations stand out from Tencent's announcement: the model participated in auto-optimizing its own training methods, data strategy, evaluation system, and low-level operators — Tencent's described "recursive self-improvement loop" — and it auto-analyzed inference bottlenecks to apply operator fusion and communication optimization, improving end-to-end throughput by 31.8% versus baseline across context lengths and concurrency levels (vendor-reported).
Evidence: The Blind Expert Evaluation
Hy4 preview's headline evidence is Tencent's blind evaluation: 163 experts rated 203 engineering tasks, scoring Hy4 preview 2.99/4 versus GLM-5.3's 2.92 (46.8% wins, 12.8% ties, 40.4% losses against GLM-5.3) and Kimi K3's 2.94 (51.2% wins, 7.9% ties, 40.9% losses). This is a vendor-run evaluation — the experts and rubric are Tencent's — and it is a preference-style signal, not a standard benchmark. It is also, for now, the only numeric quality evidence: the repository's benchmark section presents results as chart images, and Tencent's newsroom did not publish tabular coding/math/reasoning scores at launch.
What the expert result does and doesn't tell you: it suggests experienced engineers modestly preferred Hy4 preview over the current Chinese open flagships on real engineering tasks — a meaningful signal for a coding audience — but without public numeric scores on standard suites (Terminal-Bench, DeepSWE, etc.), cross-model comparison with GLM-5.3-Flash or DeepSeek V4 Pro remains indirect. See our Hy4 vs GLM-5.3 coding comparison for the full treatment.
Coding & Engineering Performance
For coding specifically, the evidence is: (1) the blind expert evaluation above (engineering tasks), (2) Tencent's product positioning in WorkBuddy/CodeBuddy — long-horizon task understanding, planning, debugging, verification, and better front-end visual quality — and (3) a known limitation the repo itself discloses: early versions of the model can "spend longer than necessary reasoning" and tend to "over-verify" their own work, which can make it slower and more verbose than users expect on straightforward tasks. There is no public Terminal-Bench or DeepSWE number for Hy4 preview as of August 30, 2026. The over-reasoning trait is especially relevant for coding agents, where extra verification loops burn tokens; Tencent's recommended sampling is temperature=0.9, top_p=1.0 with reasoning defaulting to "high" (deep chain-of-thought), and direct replies available via reasoning_effort: "no_think".
Context & Long-Context Work
Hy4 preview carries a 1M+ token context window per Tencent's newsroom, supported by the compressed Gated DSA attention and IndexCache. Combined with the 770B backbone, this targets whole-repository coding, long financial documents, and research workloads — the same use class as GLM-5.3-Flash and Qwen3.8-Flash-Next. Caveat: at ¥18 per million output tokens (≈$2.54), long-context agent loops are roughly 5x more expensive than GLM-5.3-Flash's launch pricing, so 1M-token workloads should be cost-modeled before committing. Our long-context models ranking compares the trade-offs.
Pricing & Availability
Pricing
Per Tencent's official announcement (accessed August 30, 2026): ¥6 per million input tokens, ¥18 per million output tokens, cache hits from ¥0.3 per million tokens (≈$0.85 / $2.54 / $0.04 at ~7.1 CNY/USD — our conversion, marked as estimate). WorkBuddy and CodeBuddy offered a two-week free trial at launch, and Hy3's free period was extended to September 30. This is the most expensive of the three August open releases by a wide margin — roughly 5x GLM-5.3-Flash's list input price and ~2.2x Qwen3.8-Flash's converted input price.
Where you can use it
- Hosted products: WorkBuddy / CodeBuddy (domestic and international), Yuanbao, and ima — first release simultaneous across Tencent's product surface
- API: Tencent Cloud Tokenhub and OpenRouter
- Self-hosting: Apache-2.0 weights on Hugging Face, ModelScope, GitCode, and CNB; official vLLM image (
vllm/vllm-openai:hy4-preview, e.g.--tensor-parallel-size 8 --attention-backend FLASHMLA_SPARSEwith MTP speculative decoding) or SGLang image (lmsysorg/sglang:hy4-preview,--tp-size 8, NEXTN); OpenAI-compatible endpoint after self-hosting (base_url="http://127.0.0.1:8000/v1"); finetuning pipeline in the repo; quantization via AngelSlim
Practical Use Case Analysis
- Whole-repo coding agents (self-hosted): the strongest fit — 1M+ context, Apache-2.0, and expert-preference evidence on engineering tasks; budget for the over-reasoning trait and pin
reasoning_effortwhere possible - Office & financial document analysis: Tencent's announced focus (complex office environments, financial analysis, cross-file collaboration, full document/spreadsheet/presentation delivery) — pairs with the 1M context, but no published benchmarks yet to verify claims
- Game development prototyping: one-sentence requirements → playable prototype per Tencent's positioning; treat as vendor claim until independently tested
- Scientific research: reasoning for AI R&D, molecular dynamics, condensed-matter physics, basic math — again vendor-positioned, not yet benchmark-verified
- API production workloads: viable but priced at a premium vs this week's alternatives — model the cost before choosing it over GLM-5.3-Flash pricing or Qwen3.8-Flash
Who Should Use Hy4 Preview?
- Self-hosting teams wanting the most capable Apache-2.0 model available — no other open flagship this week combines 770B scale, 49B active, and permissive licensing
- Engineering teams that trust expert-preference signals — the blind eval edge over GLM-5.3/Kimi K3 is the strongest launch evidence
- Tencent ecosystem users — WorkBuddy/CodeBuddy/Yuanbao/ima got day-one access plus a two-week free trial
- Researchers — the recursive self-improvement pipeline and AngelSlim quantization tooling are novel to study
Who Should Wait
- Teams that require public benchmark tables — none were published at launch; GLM-5.3-Flash and DeepSeek V4 Pro ship with them
- Latency-sensitive production — the disclosed over-reasoning/over-verification behavior adds token cost and latency on simple tasks
- Budget-constrained API consumers — ¥18/M output is the week's most expensive open-model rate
Pros and Cons
Pros
- Most capable-spec open model of the week: 770B total / 49B active, 1M+ context, Apache-2.0
- Blind expert eval edges GLM-5.3 (2.99 vs 2.92) and Kimi K3 (2.94) on engineering tasks (vendor-run)
- First-class open release: HF/ModelScope/GitCode/CNB, official vLLM + SGLang images, finetuning and quantization tooling
- Native MTP speculative decoding and a vendor-reported 31.8% throughput gain from auto-optimized kernels
- Day-one access across WorkBuddy, CodeBuddy, Yuanbao, and ima, with a two-week free trial
Cons
- No public numeric benchmark scores at launch — charts in the repo, no tables
- "Preview" label: Tencent itself signals early-stage release; full Hy4 models still to come
- Known over-reasoning / over-verification behavior adds latency and token cost
- Most expensive API of the week's open releases (¥6/¥18, ≈$0.85/$2.54 per million)
- All quality evidence vendor-run (blind eval) or vendor-positioned (product demos)
Hy4 Preview vs. This Week's Open Releases
Hy4 preview completes a three-release open week: GLM-5.3-Flash (Aug 26, 320B/18B, MIT, $0.15/$0.50 with full benchmark tables), Qwen3.8-Flash-Next (Aug 26, 125B/6B, Apache-2.0, 262K native), and Hy4 preview (Aug 28, 770B/49B, Apache-2.0, 1M+). Hy4 is the scale and licensing play with the thinnest evidence; GLM is the documented value play; Qwen is the efficiency preview. For the coding-focused Hy4 vs GLM-5.3 question, see Tencent Hy4 vs GLM-5.3, and for rankings, Best Open-Weight AI Models for Coding 2026.
Final Verdict
Hy4 preview is the most ambitious open release of August 2026 — Tencent's flagship-scale reasoning model, permissively licensed, with a real 1M+ context and an expert-preference signal that beats the current Chinese open leaders. It is also the least evidenced: no public benchmark tables, a disclosed over-reasoning tendency, and a name that says preview. Our independent assessment: self-host it if you want the most capable Apache-2.0 weights available this week and can tolerate early-release roughness; wait for the full Hy4 line or the promised benchmark data before betting production API spend on it — and compare the cost against GLM-5.3-Flash before you commit to either.
FAQ
What is Tencent Hy4 preview?
Tencent's next-generation MoE flagship, released and open-sourced August 28, 2026: 770B total / 49B active, 1M+ context, Apache-2.0, positioned for AI coding, office/finance, game dev, and research.
Is Hy4 preview open source?
Yes — Apache-2.0 weights on Hugging Face, ModelScope, GitCode, and CNB, with official vLLM and SGLang deployment images.
How good is Hy4 preview at coding?
No public numeric coding benchmarks at launch. Tencent's blind eval (163 experts, 203 tasks) scored it 2.99/4 vs GLM-5.3's 2.92 and Kimi K3's 2.94 — vendor-run, preference-style evidence.
How much does Hy4 preview cost?
¥6/M input, ¥18/M output, cache hits from ¥0.3/M (≈$0.85/$2.54/$0.04 at ~7.1 CNY/USD). Two-week free trial on WorkBuddy/CodeBuddy at launch.
Where can I access Hy4 preview?
WorkBuddy/CodeBuddy/Yuanbao/ima products, Tencent Cloud Tokenhub and OpenRouter APIs, or self-hosted from Apache-2.0 weights.
What are Hy4 preview's limitations?
Over-reasoning/over-verification on simple tasks (per Tencent's repo), no public numeric benchmarks at launch, and the explicit preview status of the release.
Sources (accessed August 30, 2026): Tencent official newsroom release "Tencent Releases and Open-Sources Tencent Hy4 preview," Tencent-Hunyuan/Hy4-preview GitHub repository (architecture, license, deployment, blind eval), Tencent Cloud developer coverage of the announcement, and Baidu Baike chronology. Vendor-reported figures are reproduced as published; CNY→USD conversions approximate at ~7.1 CNY/USD. ToolStep claims no laboratory test results. Learn more about our editorial policy.