
Hy4 flash vs DeepSeek V4.1 Flash: Tencent Sendiri yang Memilih Pertarungan Ini
- GoogleBARUGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1 juta token · 90 tok/s
- AnthropicBARUAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1 juta token · 96 tok/s
- orcaBARUOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaBARUOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1 juta token · 87 tok/s
- OrcaBARUOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1 juta token · 133 tok/s
- OrcaBARUOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1 juta token · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1 juta token · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1 juta token · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1 juta token · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1 juta token · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Kecerdasan
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1 juta token · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Kecerdasan
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1 juta token · 150 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Kecerdasan
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Kecerdasan
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1 juta token · 120 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Kecerdasan
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Kecerdasan
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1 juta token · 219 tok/s
Tencent chose the opponent. The research post for Hy4 flash, published on 16 November 2026, prices its own per-task cost against exactly one rival model — DeepSeek V4.1 Flash, in its 1017 snapshot, at the maximum effort tier and at DeepSeek's busy-hour rates — and reports one-eighth. That is an unusually specific comparison for a launch post, and it tells you where Tencent thinks the fight is: not against Alibaba's flagship, not against its own Tencent HY4, but against the cheap open-weight model that coding agents have been running on since September. Three weeks of published numbers sit on DeepSeek's side of that comparison. One chart sits on Tencent's.
Jadi, pertanyaan yang berguna bukanlah model mana yang lebih baik secara abstrak. Pertanyaannya adalah apakah keunggulan harga input sebesar 2,1 kali lipat tetap bertahan saat dihadapkan pada hal-hal yang berbeda antara keduanya, dan apakah angka seperdelapan per tugas berarti seperti yang akan diasumsikan pembaca. Keduanya dapat dijawab dari materi yang telah dipublikasikan, dan jawabannya mengarah ke arah yang berlawanan.
Apa yang sebenarnya dirilis oleh masing-masing perusahaan
Mulailah dengan artefak, karena artefak-artefak itu tidak setara dan perbedaannya menjelaskan sebagian besar ketidaksepakatan di hilir.
DeepSeek V4.1 Flash dirilis pada 10 September 2026 di bawah lisensi MIT: 48 shard safetensors dengan total sekitar 510 gigabyte, laporan teknis berjudul "Pushing the Limits of KV Cache Compression", kartu model dengan tag deployment, varian fp8, dan API dengan kartu off-peak dan peak yang dipublikasikan. Snapshot yang diperbarui, 1017, menggantikannya pada 17 Oktober dan kini dilayani dengan nama API yang sama. Repositori ini memiliki 1,69 juta unduhan dan 4.978 suka. Artificial Analysis memberinya skor 32,6 pada Intelligence Index dengan upaya maksimum; Capabilities Index dari Epoch AI menempatkan checkpoint 1017 pada 159, peringkat ke-19 dari 253 model yang dilacaknya.
Hy4 flash went up on 16 November 2026 as a research post with no model card attached: 224 billion parameters total, 64 billion of them Engram parameters, 12 billion active per token, native multimodality, a 1,048,576-token context, and a plan to open-source later. No weights, no licence, no technical report, no independent score, and benchmark results delivered as PNG charts. Artificial Analysis carries Tencent HY4 at 34.6 as of 5 November and has no Hy4 flash entry at all.
Letakkan keduanya berdampingan dan perbandingannya punya bentuk yang janggal. Model yang lebih baru adalah model yang tidak dapat Anda periksa, sedangkan model yang lebih lama adalah model yang dapat Anda unduh, fine-tune, dan jalankan di perangkat keras Anda sendiri. Asimetri itu merasuki setiap bagian di bawah ini.
Kartu harga, dibandingkan secara jujur
Kedua vendor menerbitkan tarif daftar, dan keduanya menerbitkan lebih dari satu tarif, dan di sinilah sebagian besar perbandingan menjadi keliru.
• Input, per juta token — Hy4 flash $0.070 flat; DeepSeek V4.1 Flash $0.15 di luar jam sibuk, $0.30 selama jam sibuk.
• Output, per million tokens — Hy4 flash $0.281; DeepSeek $0.60 off-peak, $1.20 peak.
• Cache hit, per million tokens — Hy4 flash $0.003; DeepSeek $0.003 off-peak, $0.006 peak.
• Peak definition — DeepSeek's peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is off-peak, which is to say that for a US or European team the cheap card is the one in force for most of its working day.
• Konteks — Hy4 flash 1.048.576 token; DeepSeek V4.1 Flash 1.000.000 token dengan batas keluaran 384.000 token.
• Kartu Tiongkok Daratan — Hy4 flash ¥0.5 / ¥2 / ¥0.02 per juta token, yang menjadi sumber angka dolar tersebut, bukan dari tarif internasional yang diterbitkan secara terpisah.
Read at the same hour, the gap is 2.1 times on input and 2.1 times on output: $0.070 against $0.15, $0.281 against $0.60. During DeepSeek's peak windows it widens to 4.3 times. On cached input it is a wash off-peak, because both land at $0.003 — and that matters more than the headline, since a coding agent that re-sends a system prompt, a tool schema and a file tree on every step is paying the cache rate most of the time. On the workload shape this matchup is actually about, the two models are priced identically.
That last line is the one that defuses most of the argument. Tencent's one-eighth is not a cache-heavy comparison; it is a per-task cost on WorkIndex, which folds in output length, retry rate and step count alongside the rate card. Two of those three inputs are things Tencent controls through its own post-training, and it reports the control as a single percentage: output length down 50%.

Mengapa 12 miliar parameter aktif adalah inti dari argumen arsitektur
DeepSeek V4.1 Flash is a 552-billion-parameter backbone with 16 billion active per token, built on what DeepSeek calls a Causal Encoder-Decoder design: 40 layers split evenly into a 20-layer causal encoder and a 20-layer decoder, with the technical report's emphasis on KV cache compression rather than on perception. Hy4 flash activates 12 billion per token against a 224-billion total, which is three-quarters of DeepSeek's active count and 40% of its total size.
Tencent's account of where the difference comes from is specific about mechanism and silent about structure. Hybrid attention, LatentMoE and the Engram parameters produce a KV cache one twenty-fourth the size of Hy3's and one sixth of Tencent HY4's. Self-distillation from Tencent HY4, then reinforcement learning, cut end-to-end output length by half. A megakernel with four-bit weights and a DFly decoding path keeps decode above 150 tokens per second. What is not published is anything you could falsify from a checkpoint: layer count, expert count, routing width, vision encoder size, and whether multimodality is native or attached.
The Engram parameters deserve a sentence of their own, because 64 billion of them inside a 224-billion-parameter model is not a normal arrangement. Engram, in Tencent's usage, is a memory-style component holding learned associations rather than a routed expert that runs in every forward pass. The practical consequence is that 29% of the model's total size does not get charged against per-token compute, which is a legitimate way to buy capability per parameter and an illegitimate way to compare parameter counts. A 224B model with 64B of Engram is not smaller than a 552B model in the way the two numbers suggest; it is differently shaped.
DeepSeek's design is checkable and Tencent's is not, and that is not a small difference in a head-to-head. Anyone can download 510 gigabytes of weights and count the layers. Nobody can download Hy4 flash and do anything at all.
Seperdelapan, ditelusuri
Tencent's per-task comparison is worth following closely because it is the only head-to-head number either company has published for this pair, and it is more careful than launch-post arithmetic usually is.
On WorkIndex — Tencent's internal eight-suite set covering slide creation, slide revision, human-machine co-writing, Markdown co-writing, document editing, table repair, audio and video processing, and visual understanding — Tencent reports Hy4 flash at one-thirteenth the per-task cost of Hy4 preview, one-twentieth of Tencent HY4, and one-eighth of DeepSeek-V4.1-Flash-1017 at maximum effort tier and busy-hour price. Every qualifier in that sentence is Tencent's, and the last two are the ones to hold onto: maximum effort means DeepSeek's thinking budget was set high, and busy-hour means the $0.30/$1.20 card rather than the $0.15/$0.60 one.
Take the qualifiers away and the arithmetic changes a lot. At off-peak rates with the same output lengths, the per-token gap is 2.1 times, not 8. If Hy4 flash also emits half the output tokens, per-task cost roughly halves again on top of the price difference, which puts the two models somewhere between 4 and 5 times apart on a task Tencent designed — and DeepSeek's own published behaviour, controlled thinking effort and a 384,000-token output ceiling, suggests its maximum tier is doing considerably more work per task than a model tuned to answer in half the tokens.
The production numbers Tencent volunteered are the more interesting half of its release, because they come with counters attached rather than a single score. In WeChat Pay's agent flows, against DeepSeek-V4.1-Flash-1017, Tencent reports parity or better across four process stages with about 6% fewer execution steps and 29% shorter single outputs. In WeChat Pay's static scanning, 37% less time to complete a scan and 46% fewer tokens at equal detection quality. In Honor of Kings root-cause analysis, 22% to 39% fewer requests and 40.5% to 61.2% fewer tokens than DeepSeek and GLM-5.3-Flash — while making 51.2% and 7.7% more tool calls than each. In Marvis, task satisfaction up 6.2% on app installation, 4.4% on system setup and diagnostics, 6.9% on document processing and research reports.
Two things are visible in that list. The token and step reductions are large and consistent, which is what a 12B-active model trained to shorten its answers should produce. And the tool-call increase is the honest footnote: Hy4 flash gets there with more, shorter round trips. On a per-request billing model that is a wash or a win; on a per-step wall-clock budget it is a different trade, and Tencent's own framing — that it is 88% faster in WeCom — is a latency claim rather than a work claim.
Asimetri bukti, yang merupakan keputusan sebenarnya
Berikut adalah apa yang dapat disampaikan kedua model tentang diri mereka sendiri.
DeepSeek V4.1 Flash memiliki skor harness publik, laporan teknis yang bisa dibaca, bobot yang bisa dijalankan, dan tiga bulan tumpukan inferensi yang berinteraksi dengan model ini: kuantisasi, resep penyajian, dan serangkaian keanehan terdokumentasi yang sudah dialami orang lain. Posisi tolok ukurnya tidak spektakuler — 90,6 pada Terminal-Bench 2.1, 74,2 pada DeepSWE v1.1, 65,4 pada NL2Repo-Bench saat rilis, dan 32,6 Intelligence Index — tetapi itu posisi yang diukur oleh orang lain.
Hy4 flash has a vendor chart, a vendor blind test of 85 employees, a vendor internal benchmark, and a sentence about future open weights. The claims are plausible and internally consistent: a KV cache a sixth the size of the flagship's, output length halved, decode above 150 tokens per second. None of them has been reproduced, and the one that would be easiest to check — whether 12 billion active parameters really does office work as well as DeepSeek's 16 billion — is exactly the one nobody outside Tencent can run, because there are no weights.
That asymmetry does not make Hy4 flash worse. It makes it unproven, and the difference between worse and unproven is where evaluation budget should go. A team that needs a cheap multimodal model it can fine-tune, self-host or audit should stop reading here: DeepSeek V4.1 Flash is the answer, and the 2.1-fold price difference is the cost of that option. A team whose workload is documents, slides, extraction and short agent loops, and which pays per token rather than per task, has a real reason to look at the Tencent card — a reason strong enough that the missing evidence is worth buying around rather than waiting for.
Menjalankan perbandingan pada satu kunci
Neither Hy4 flash nor Tencent's own roadmap makes this easy: as of 16 November, Hy4 flash is served through Tencent Cloud's TokenHub and Tencent's consumer and enterprise products, with third-party platforms announced but undated, and DeepSeek V4.1 Flash is on DeepSeek's own API and on open weights.
The half of this matchup you can run today is the half that sets the bar. deepseek/deepseek-v4.1-flash is on OrcaRouter at $0.15 in and $0.60 out per million tokens with cache reads at $0.003 — DeepSeek's own off-peak list price with nothing added — over a 1,048,576-token context, and the 1017 snapshot sits at the same rate under its own identifier so you can pin the checkpoint Tencent benchmarked against rather than the alias. The Tencent route on the same key is tencent/hy4-preview at $0.834 and $2.501, which is Hy4 flash's teacher; running your own task set on preview and on DeepSeek gives you the teacher-to-rival ratio now, and the flash tier can be slotted into the same harness when it lands.

That is the pattern worth copying from Tencent's own release. Its argument is a per-task cost on eight task types it chose. Yours should be a per-task cost on the tasks you actually pay for, measured on both models, with tokens, steps and wall-clock recorded — and the three of those together are what will tell you whether a 2.1-fold rate advantage turns into anything at the invoice.
Automatic failover is the unglamorous part of that plan. A model in its first week, served from a promotion with no stated end date, is exactly the model you want a fallback chain behind rather than a migration onto; the same key already reaches DeepSeek, so a rule that retries the incumbent before the response starts costs a config line and removes the launch-week risk from the experiment.

Apa yang bisa menyelesaikannya
Tiga hal, diurutkan berdasarkan seberapa besar hal itu akan mengubah jawabannya.
An independent score for Hy4 flash. Artificial Analysis indexed Tencent HY4 eight days after it shipped; a flash entry would at least put both models on one harness, and the 32.6-to-34.6 gap between DeepSeek's flash and Tencent's flagship is the band this matchup is arguing about. Nobody should be surprised if a distilled model lands below its teacher, and the size of that drop is the single most useful unknown in this comparison.
Weights, or a date for them. Tencent's own August preview shipped Apache 2.0 with vLLM and SGLang recipes on day one, so the company knows how to do this. If Hy4 flash follows, its efficiency claims become checkable and the 510-gigabyte download stops being the only option in the room.
A neutral per-task benchmark. Tencent's one-eighth is a good-faith number with two qualifiers attached and a harness nobody else can run. Reproduce it on public agent tasks and the argument is over in an afternoon, one way or the other.
Until then the honest summary is narrower than either vendor's: DeepSeek V4.1 Flash is the cheap multimodal model with an audit trail, Hy4 flash is the cheaper one without, and the price difference is smaller on cache-heavy work than on a rate card. That is not a verdict. It is an accurate description of what 16 November actually changed, which was a rate card and a plan — and a fight Tencent chose on a scoreboard it drew itself.
Dibandingkan dalam artikel ini4
Terdeteksi dari artikel ini · Benchmark: Artificial Analysis · diperbarui setiap hari
