
Hy4 flash vs Tencent HY4: One Family, Two Tiers, and a Twelvefold Price Gap
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 150 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 119 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
Tencent HY4 went into general availability on 4 November 2026 as a 770-billion-parameter model with 49 billion active per token, native vision, computer use and a $0.834 input / $2.501 output price per million tokens. Twelve days later, on 16 November, Tencent shipped Hy4 flash: 224 billion total parameters with 12 billion active, the same one-million-token context, and $0.070 input / $0.281 output. Hy4 flash was built by distilling Tencent HY4 and then training with reinforcement learning to talk less, and Tencent's own guidance says as much — use the flash model for deep search, documents, slides, frontend work and computer use, and keep the flagship for long-horizon software engineering. That guidance is unusually candid for a launch, and it is also the most useful thing in either release.
So this is not a contest between rivals. It is a company splitting its own family in two and telling you which half to buy, and the interesting questions are how much of the flagship survives in the small model, where the split actually falls, and what the twelvefold price difference buys besides tokens.
Three releases in eleven weeks
The lineage explains most of the price. Tencent has shipped this family three times since August, and each release moved a different variable.
• 28 August 2026 — Hy4 preview. Apache 2.0, weights on Hugging Face, ModelScope, GitCode and CNB, 770B total / 49B active, 78 layers of which one dense and 77 mixture-of-experts with 256 routed experts plus one shared, top-8 routing, a 1,048,576-token context, a 10B multi-token-prediction layer with 0.7B active. ¥6 / ¥18 / ¥0.3 per million tokens. Eight published benchmark rows, vLLM and SGLang recipes, published sampling defaults.
• 4 November 2026 — Tencent HY4. Same backbone scale, substantially more pre-training and post-training, and two new capabilities: native vision over fine-grained images, charts and diagrams, and computer use through mouse, keyboard and command line with specific training for Blender, CAD and long-horizon cross-application work. Same ¥6 / ¥18 / ¥0.3 price. No weights, no licence, no model card, no benchmark appendix.
• 16 November 2026 — Hy4 flash. 224B total with 64B of Engram parameters and 12B active, multimodal, 1,048,576-token context, ¥0.5 / ¥2 / ¥0.02. No weights, no licence, no model card, benchmark numbers published as chart images only.
The direction of travel is visible: the family has been getting cheaper and less open at the same time. The cheapest member is also the one with the least paperwork, and the only version anyone can download is the one from August.
What the flagship still does that flash does not

Two capabilities sit in the flagship and are not claimed for the small model in the same terms.
Vision at scale. The HY4 release demonstrated a 2.8-gigapixel aerial photograph in which the model located a two-metre roof component, which is less a vision result than a demonstration that a million-token window can hold a tiled image and a query at once. Hy4 flash is called multimodal, and one of the eight task types in its WorkIndex suite is visual understanding, but the release does not say whether the encoder is its own or bolted on, how large it is, or whether video goes in at all. For dense-document and chart work the flagship is the model with a published demonstration behind it.
Computer use and long-horizon desktop work. Tencent trained HY4 specifically for driving a desktop across applications, ran it in a sandbox built on its own Agent Runtime with several hundred applications installed, and named Blender and CAD as target workloads. Hy4 flash is recommended for computer use too, which is a meaningful concession — but the flagship is the one with the training description attached, and the internal WorkBuddy data shows flash giving up ground on long coding tasks even while reaching parity on daily work.
Everything else in the flagship's column is scale rather than capability: 49 billion active parameters against 12 billion, 78 documented layers against an undisclosed count, 256 routed experts against an undisclosed number. More active parameters buy depth on hard reasoning and long tool chains. They do not buy access to the context window, which is the same on both, or to multimodality in principle, which both claim.
Where the twelvefold difference comes from
The price gap is not a discount. It is four times fewer active parameters plus a set of serving changes, and Tencent is unusually specific about the second part.
The efficiency work, per the vendor: hybrid attention, LatentMoE and the Engram parameters together produce a KV cache one twenty-fourth the size of Hy3's and one sixth of Tencent HY4's. Self-distillation from HY4 followed by reinforcement learning cut end-to-end output length by 50%. An inference megakernel with four-bit quantization and a decoding path called DFly sustains more than 150 tokens per second. Each is plausible, each is consistent with the architecture, and none has been independently reproduced.
The KV cache figure is the one that changes deployment arithmetic rather than just price. On a million-token context, cache size decides how many concurrent long requests fit on a node, and a sixth of the flagship's cache at a twelfth of the price is a combination that can turn an unprofitable workload into a profitable one without any change to model quality. The 50% output-length cut is the one that changes quality, and it is the number to test: shorter answers are cheaper because they are shorter, and whether they are shorter because the model is efficient or because it stopped thinking is exactly what a vendor cannot tell you about its own model.
For reference on the family's own trajectory, Tencent reports per-task cost on its internal WorkIndex suite at one-thirteenth of Hy4 preview and one-twentieth of Tencent HY4. Per-task is a different measurement from per-token, folding in output length and retries, so the twentieth is the honest operational figure and the twelfth is the rate card. Both are large enough that the direction is not in doubt.
The distillation question nobody has answered
A distilled model inherits its teacher's behaviour on the tasks the distillation covered and invents its own on the rest. Tencent says Hy4 flash was self-distilled from HY4, which makes the flagship both the competitor and the training signal — an arrangement that usually produces a small model that is very good at what the teacher was drilled on and noticeably weaker elsewhere.
The company's internal product data is consistent with that pattern. In WorkBuddy, Hy4 flash runs at roughly one-fifteenth the cost of Tencent HY4 with task time down 70%, parity on daily tasks, and a small drop on long coding. In 微信支付 agent flows it matches or beats DeepSeek-V4.1-Flash-1017 with 6% fewer steps and 29% fewer output tokens. In 王者荣耀 root-cause analysis it lands close to both DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash on localisation while using 22% to 39% fewer requests and 40% to 61% fewer tokens. In 企业微信 it produces 88% faster output with 60% fewer thinking tokens than HY4.
Read together, those numbers describe a model that is excellent at the shape of task the family was trained on — documents, agents, structured output, retrieval — and that gives back a little on the tasks where the teacher's extra depth was doing the work. That is the expected outcome of distillation and it is not a criticism; it is the reason to keep both tiers rather than replace one with the other.
What neither model has
Neither the flagship nor the flash release comes with weights. Tencent says it plans to open-source Hy4 flash and to iterate frequently, but no licence, no date and no repository exist. Tencent HY4's own weight release is likewise announced as a plan. The only member of this family you can download today is Hy4 preview, Apache 2.0, on Hugging Face since 28 August, and that remains true eleven weeks and two releases later.
Neither has an independent benchmark score. Artificial Analysis has published nothing for Tencent HY4 or Hy4 flash. Every figure in both releases is Tencent's, the benchmark tables exist as chart images rather than text, and the blind test behind the flash model's headline claim is 85 employees of Tencent using a Tencent product on tasks Tencent selected. That does not make the numbers wrong. It makes them a starting hypothesis.
Neither has a published sampling configuration, a system card, or a statement about refusal behaviour. The preview shipped all three.
The practical split
Tencent's guidance maps onto workload shape more cleanly than most vendor advice, and it is worth restating in operational terms.
Default to Hy4 flash for anything that is mostly reading. Deep search, document processing, extraction over long inputs, slide generation, simple frontend scaffolding, and agent loops that spend most of their tokens re-sending context. The tenfold cache discount and the smaller KV footprint make long-context retrieval the workload where the cheap tier is most clearly correct, and the internal data covers exactly that ground.
Reach for Tencent HY4 when the task is a long chain of decisions. Multi-file software engineering, long-horizon desktop automation, CAD and Blender work, and anything where a run has to hold a plan together across dozens of steps. This is the case Tencent itself concedes, and the flagship's extra 37 billion active parameters are the mechanism.
Use Hy4 preview as the control. It is the same family, the same 770B/49B scale as the flagship, a published benchmark appendix, and Apache 2.0 weights you can run yourself. When you are deciding whether flash is good enough, the preview gives you a reference point that does not depend on trusting a chart.
Where to run the comparison

Neither Tencent HY4 nor Hy4 flash is routed through OrcaRouter. What is here is tencent/hy4-preview at the provider's list price of $0.834 per million input and $2.501 per million output, plus a rate-limited free tier, and tencent/hy3 at $0.18 and $0.59 for the older generation. The two models in this article are reachable through Tencent's own TokenHub, where the flash tier is currently running a 50%-off promotion for an unstated period and WorkBuddy is free until 30 November.
The preview route is worth more than it looks. It runs the same architecture at the same scale as the flagship, publishes an eight-row benchmark appendix you can re-run against the checkpoint, and costs a quarter of what evaluating the flagship would. If your question is whether this family's large model is good enough for your work, the preview answers it at a fraction of the price; if the answer is yes, Hy4 flash is the version to price against it, and the distilled relationship means the comparison is unusually clean — same family, same context, same tokenizer lineage, twelvefold difference in what you pay.
One reason to keep the evaluation on a router even for a family you can only reach from its vendor: a model released six days ago has no public uptime record, and a swap to tencent/hy4-preview on the same key is a one-line change if an endpoint misbehaves during a launch-week promotion. Nothing is added on top of what Tencent charges — the list price is the price — and when Tencent adjusts a rate the new number is what you pay on the next call, not on the next release of somebody's comparison sheet. The same credential also opens more than two hundred other models if the evaluation widens beyond this family.

The tier split is the announcement
The most consequential thing Tencent said on 16 November was not the price. It was that the flagship stays for long-horizon software engineering while the cheap model takes everyday work — a company dividing its own product line along the axis its customers actually feel, and conceding that the cheap model is not a replacement for the expensive one.
For anyone inside that split, the decision is easy and the numbers are large: a twelfth of the input price, a ninth of the output, a hundredth of the cache, four times fewer active parameters, and a smaller KV footprint on the same million-token window. For anyone outside it — long agentic runs, desktop automation, multi-file code — the flagship is still the family's answer and its own vendor says so.
What is missing is proof, and it is missing from both tiers. No independent score, no weights for either new model, benchmark tables you have to read as pictures, and a headline comparison drawn from an internal panel. A twelvefold price difference is a strong enough claim that it will be tested by people outside Tencent within weeks. Until it is, the sensible position is the one Tencent's own guidance implies: run the cheap tier on the tasks it was built for, keep the flagship where the work is long, and measure both rather than choosing on the rate card.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
