
Hy4 flash vs Qwen3.8-Flash: Twice the Active Parameters, Half the Price per Token
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 138 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 175 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 86 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 120 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 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 · 44 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 · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 116 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
The cheaper model here is the bigger one. Hy4 flash fires 11.3 billion text parameters on every token and lists at $0.070 per million input tokens. Qwen3.8-Flash fires roughly 6 billion and lists at $0.15. Nearly twice the active compute per token, less than half the input price — and neither of those two numbers is the one that decides which model you should actually call.
Hy4 flash is two days old. Tencent put it on its own API on 16 November 2026, published the weights the next day, and shipped three repositories for it: the instruct checkpoint, an FP8 build and a base model. It is a 220-billion-parameter mixture of experts with 12 billion active, a 1,048,576-token context, image input and text output, under the Tencent Hy License. Qwen3.8-Flash has been on QwenCloud since Alibaba opened it alongside the Qwen3.8-Flash-Next weights on 26 August 2026: a 125-billion-parameter expert body activating about 6 billion, plus a 51-billion-parameter n-gram embedding table, taking text, image and video in and returning text, at 1M context and a 131K maximum output.
So the shape of this comparison is unusual. One model is the best-documented cheap tier in the market and the other is the least documented. One has an independent benchmark row and the other has a 404. And the two price cards are denominated in different currencies, sold into different markets, and hedged by different amounts of fine print. The interesting part is not the 2.1x on the input line — it is what happens to that gap once you stop pricing tokens and start pricing finished work.
The inversion on the rate card, before anything else
Tencent's mainland card for Hy4 flash reads ¥0.5 per million input, ¥2 per million output and ¥0.02 per million cached input. Tencent quotes the same three lines in dollars as $0.070, $0.281 and $0.003. QwenCloud's card for Qwen3.8-Flash reads $0.15 in, $0.47 out, $0.016 for a cache read and $0.20 per million for explicitly creating a five-minute cache.
• Input — $0.070 against $0.15. Hy4 flash is 2.1x cheaper.
• Output — $0.281 against $0.47. Hy4 flash is 1.7x cheaper.
• Cached input — $0.003 against $0.016. Hy4 flash is 5.3x cheaper on the line item that dominates any agent re-sending a fixed prompt.
• Blended at 3:1 input-to-output — about $0.123 per million against $0.230.
• Blended at 1:3, output-dominant — about $0.228 against $0.390.
That is a real gap and a smaller one than the headline ratio implies, because the output column is where both models actually spend money on agentic work and the output gap is only 1.7x. It is also smaller than the gap either of these models shows against a Western flagship, which is the company each of them keeps in this tier.
The honest caveat on all five lines: these are not the same kind of price. Tencent's is a mainland RMB list rate that Tencent has converted for English-language readers, and Tencent has already run promotional discounts on this model in its first week. QwenCloud's is a USD list rate on a global endpoint. A team buying in dollars from outside mainland China should treat the Hy4 flash figures as the floor of what is available rather than the number they will be invoiced at, and should re-read the card before modelling anything on it.
Run the volume anyway, because the shape holds. Ten million input tokens and two million output tokens a day costs $1.26 on Hy4 flash and $2.44 on Qwen3.8-Flash — $461 against $891 over a year, a difference of about $430 at a workload size a small team can actually generate. Push the mix toward cache, as any multi-turn agent does, and the gap widens: two million fresh input, eight million cache reads and two million output comes to $0.73 a day against $1.37, because the cache-read ratio is 5.3x rather than 2.1x.

Both vendors spent nearly a third of the parameters on the same idea
This is the part of the matchup that neither vendor's launch post puts next to the other's, and it is the reason the two models are priced within a factor of two instead of a factor of ten.
Hy4 flash carries 63.5 billion Engram parameters — learned bigram and trigram representations, retrieved through multi-head hash lookups, plugged into the backbone at layers 1, 11, 21 and 31. That is 28.9% of the model's 220 billion parameters sitting in a lookup table rather than in experts. Qwen3.8-Flash carries a 51-billion-parameter n-gram embedding table keyed on the current token plus several preceding ones, out of roughly 180 billion parameters on disk — 28.3%.
Two labs, two generations, the same conclusion: the cheapest way to add capability to a sparse model is to stop paying per-token compute for it. Both tables can live in host memory and be prefetched asynchronously, so neither permanently occupies GPU memory. Neither adds a matrix multiply to the forward pass. Both are, in effect, a bet that a large fraction of what a language model needs to know is locally predictable from the last few tokens, and that the sparse expert body should be reserved for what is not.
The attention stacks rhyme too. Hy4 flash interleaves IndexPool sparse attention with 512-token sliding windows at a 1:3 ratio across 42 layers, and Tencent reports the resulting KV cache at 1/24 of Hy3's and 1/6 of Hy4's. Qwen3.8-Flash uses Gated DeltaNet with Qwen Sparse Attention, compressing history into fixed-size states in three of every four layers and routing the rest through a block-level indexer. Both are answers to the same question — how do you serve a million-token window without paying for a million-token KV cache — and both answers land on some variety of "don't attend to everything."
The difference is what the extra active parameters buy. Hy4 flash runs 11.3 billion text parameters against Qwen3.8-Flash's 6 billion, on 41 MoE layers with 16 of 512 routed experts plus one shared expert, in a latent space of 1,536 dimensions. That is nearly double the per-token compute, in a model that charges less than half per token. Whatever else is true, Tencent is not recovering that cost from the customer — it is recovering it from the serving stack: 4-bit weights, megakernels, and DFly2, a 0.85-billion-parameter drafter that reuses the target model's sparse KV and index selections so each draft token attends to at most about 2,000 keys regardless of context length. Tencent puts decoding above 150 tokens per second with those three together.
What 29.5 minutes per task actually buys
Per-token prices are the wrong unit for both of these models, and the independent data makes that concrete on one side of the comparison.
Artificial Analysis scores Qwen3.8-Flash-Next — the open-weights artefact, not the hosted SKU — at 28.6 on Intelligence Index v5.0.1, at $0.569 per task, over 29.5 minutes per task, at 83.6 output tokens per second, with 2.5 seconds to first token and 68.3 seconds to first answer token.
Those five numbers multiply out into something more useful than any of them alone. At 83.6 tokens per second for 1,768 seconds, a task produces roughly 148,000 output tokens. At $0.47 a million that is about seven cents of output. The other fifty cents of the $0.569 is input, which at $0.15 a million implies at least 3.3 million input tokens across the task — more if part of it was served from cache at the cheaper rate. A 3.3-million-token input against a 1M context is not one prompt. It is a loop: read, act, read again, twenty or thirty turns deep, with the context rebuilt or carried forward each time.
That is the workload both models are built for, and it is why Hy4 flash's design choices read the way they do. Tencent's post-training pipeline cut average end-to-end output by 50% through reinforcement learning, and its internal blind evaluation reports a 6.7-minute median task completion at RMB 0.26 per task — about $0.037 at prevailing rates. Its internal WorkIndex suite puts Hy4 flash at 42.1 against 39.7 for Hy4 preview, at an average cost per task Tencent states as one-thirteenth of Hy4 preview, one-twentieth of Hy4 and one-eighth of DeepSeek-V4.1-Flash-1017 at peak-hour pricing.
Do not divide those figures by Artificial Analysis's. A 6.7-minute median on 259 WorkBuddy tasks across 85 internal users is not a 29.5-minute median on the Intelligence Index, and a RMB per-task figure from a vendor's own harness is not a dollar per-task figure from an independent one. The comparison they support is directional and it is worth stating plainly: Qwen3.8-Flash's measured cost per task is dominated by how many tokens it spends and how long it takes, not by what it charges per token. A model that writes half as much and finishes in a fifth of the time can be 1.7x cheaper per output token and an order of magnitude cheaper per finished job. Whether Hy4 flash actually is that model is precisely what nobody has measured yet.
The scoreboard has one row, and it is not for the model you would call
Two disclosures, both of which change how much weight the numbers above can carry.
First: Hy4 flash has no Artificial Analysis page. Not a low score, not an unranked entry — https://artificialanalysis.ai/models/hy4-flash returned 404 on 18 November 2026, three days after the API opened and a day after the weights landed. There is no independent measurement of this model at all. Tencent's published benchmark table covers the pre-trained base checkpoints, Hy4 flash Base against Hy4 Base and DeepSeek's base models, which tells you about pre-training and nothing about the instruction model or the reasoning mode you would actually call. The only instruction-model evidence is Tencent's own: WorkIndex, the blind test, and a per-task cost ratio.
Second: the 28.6 and the $0.569 belong to Qwen3.8-Flash-Next, the downloadable checkpoint. The hosted Qwen3.8-Flash on QwenCloud is a separately tuned serving build; Alibaba sells it as the productionised member of the line and the two are not interchangeable. Artificial Analysis also records Qwen3.8-Flash-Next at a 256,000-token context — the checkpoint's native window before YaRN extension — while QwenCloud serves the hosted model at 1M with a 991K maximum input. If you are pricing the hosted model, you are relying on QwenCloud's card and Alibaba's own evaluations for everything except the intelligence score, which is the checkpoint's.
So the asymmetry is stark in a way that is easy to state and hard to resolve: one model has an independent score that belongs to a different artefact, and the other has no independent score at all. Any article that hands you a winner on capability here is inventing one. What can be said is narrower and still useful. Qwen3.8-Flash at 28.6 sits below Tencent Hy4, the 770-billion-parameter flagship from the same family, at 34.6, and below Claude Haiku 5.5 at 36.5 on the same index version — a solid mid-tier score, which is what a small flash model should produce. Hy4 flash, on Tencent's own internal suite, scores above Hy4 preview and well below Hy4, which is the same relative position by a different instrument.
Video is the one row that decides itself
Everything above is arguable. This is not.
• Hy4 flash input — text and images, through a SigLIP2-So400m-derived encoder over an MLP projector. Output is text.
• Qwen3.8-Flash input — text, image and video. Output is text.
If your pipeline ingests a screen recording, a call capture or a video file, Hy4 flash cannot accept it and Qwen3.8-Flash can. The open weights do not help, an adapter does not help, and no amount of price advantage closes it. Both hold a million tokens of context, both handle documents and screenshots, and on that shared ground the comparison stays open; on video it is closed.
Tencent's own recommended workloads for Hy4 flash are consistent with that boundary — deep search, document processing, slide creation, simple frontend demos, computer use — and its own guidance sends long-horizon software engineering to Hy4 instead. That is a vendor conceding a ceiling, which is more informative than most launch copy and is worth taking at face value.
Both are downloadable, and the downloads are not equivalent

Hy4 flash shipped as open weights on 17 November: tencent/Hy4-flash, tencent/Hy4-flash-FP8 and tencent/Hy4-flash-base. A file listing on 18 November puts the BF16 checkpoint at 440.5 GB across 148 shards and the FP8 build at 287.1 GB. Qwen3.8-Flash-Next has been public since 24 August and measures 360.0 GB across 131 shards.
• Serving recipe — Tencent publishes pinned images, vllm/vllm-openai:hy4-flash and lmsysorg/sglang:hy4-flash, with the tensor-parallel-8 and speculative-decoding flags written out. Qwen3.8-Flash-Next has had SGLang, vLLM and Transformers support since its release week.
• Transformers — Hy4 flash requires an install from a pull-request ref, refs/pull/50903/head, because support is not yet on main. Qwen3.8-Flash-Next loads from main.
• Quantisation — Tencent ships an official FP8 build and an AngelSlim GGUF; the Qwen ecosystem has had four months to produce community quantisations of its own.
• Ecosystem signal — Qwen3.8-Flash-Next recorded 2,983,542 downloads in the last month and 7,276 likes on 18 November. Hy4 flash, published the day before, showed 96 likes and a zero download count.
• Hardware — 440 GB of BF16 weights does not fit on eight 80 GB cards, so the FP8 build at 287 GB across eight H100s is the realistic single-node deployment, and that is exactly the TP=8 configuration Tencent's own recipe specifies.
Pinned images and a PR-ref install are not disqualifying; they mean you are running the builds Tencent tested rather than builds that have been released, which is a difference in who owns the next bug. Qwen's advantage here is entirely a function of four months of elapsed time, and it will erode. Today it is real, and it is the second-largest gap in this comparison after the missing benchmark row.
The two licences both have a model-as-a-service clause, and they are different clauses
Both models are free to download and neither is unconditionally free to resell, but the triggers are not the same shape and the difference matters if you host models for other people.
• Tencent Hy License v1.0 — an Apache 2.0 derivative. Section 10.2 requires a separate written licence if you operate a model-as-a-service business and your worldwide gross revenue exceeds US$50 million in any twelve-month period, or your product passes 100 million monthly active users. Section 10.3 exempts internal use.
• Qwen Community License 1.0 — requires a separate licence for any Model-as-a-Service or AI Work Assistant business regardless of size, with internal use exempt. It separately requires the model name be displayed prominently if you exceed 100 million monthly active users or US$20 million in monthly revenue.
• The Qwen carve-out worth reading twice — the licence explicitly excludes "the mere relaying of requests to models hosted by other third parties" from its definition of Model as a Service. Tencent's licence has no equivalent sentence.
Neither carries a territorial restriction, which is more than can be said for several Chinese open-weight licences. If you are building an internal tool, both are free at any scale. If you are building a product that resells inference, Tencent's threshold is simpler and Qwen's relay carve-out is more permissive for a specific kind of intermediary, and which one binds you depends on which of those two businesses you are in.
What is callable today, on one key

Qwen3.8-Flash is on the OrcaRouter catalogue as qwen/qwen3.8-flash, at $0.15 per million input and $0.47 per million output with a $0.0184 cache read — QwenCloud's published prompt and completion rates passed through at 0% markup, so a vendor cut on this model reaches you the day it happens rather than at the next catalogue sync.
Hy4 flash is not on the catalogue yet. What is there for this family is tencent/hy4-preview at $0.834 and $2.501 per million with a $0.042 cache read — the model Hy4 flash was distilled from, and the right baseline to run against Tencent's own API while you wait for the flash route to land. Saying otherwise would be claiming a route we do not have.
What that asymmetry gives you is a cheap way to settle the argument yourself. Run your twenty most representative tasks on qwen/qwen3.8-flash through the key you already have, and on Hy4 flash through Tencent's. Record input tokens, output tokens, cache hits, wall-clock and whether the output was usable. The per-task ratio you get is the only version of Tencent's one-thirteenth claim that applies to your traffic, and it costs an afternoon rather than a migration.
When the flash route does land, the useful pattern is a routing rule rather than a switch: put the video-ingestion tier on the Qwen route because it is the only one of the two that accepts video, and the text-and-document tier on whichever model your own numbers favour. The routing DSL expresses that split in one endpoint, automatic failover means an unproven two-day-old route degrading does not take the pipeline with it, and the same key reaches more than 200 models without a second contract.
The decision, stated as three cases
If you need video input, the comparison ends here. Qwen3.8-Flash takes it, Hy4 flash does not, and no price argument survives a missing modality.
If you need a model with an independent score before it touches a customer-facing path, you already have your answer for the same reason. 28.6 on Intelligence Index v5.0.1 is a modest number — below the 34.6 of Tencent's own flagship and below Claude Haiku 5.5 at 36.5 — but it is measured, public and reproducible, and Hy4 flash has nothing of the kind. That will probably change; it is the single most valuable thing anyone could publish about this model.
If your workload is high-volume text and documents, latency-sensitive, and you can run your own evaluation, Hy4 flash is worth the afternoon. Nearly double the active parameters at less than half the input price is not a normal trade, the cached-input line at $0.003 makes long-running agents close to free to re-prompt, and the weights mean the price is one option rather than the only one. Go in knowing that you are the first person to measure it, that its published benchmark table describes a different checkpoint than the one you will call, and that 440 GB of BF16 weights will not fit on a single node without the FP8 build.
The number to watch is not another launch post. It is an Artificial Analysis row for Hy4 flash, because that converts a vendor's internal ratio into a figure you can set directly against Qwen3.8-Flash-Next's $0.569 per task over 29.5 minutes. If Hy4 flash lands meaningfully under that on cost per task while holding within a few index points, the price gap is the whole story and this tier reorganises around it. If it lands cheap because it does less work per task, then it is a document and slide model with an unusually good rate card, and Tencent's own deployment guidance already says so.
Compared in this article3
Detected from this article
