Tencent Hy4 vs DeepSeek V4.1 Flash: Which Cheap Open-Weight Agent Model to Pick?

Two of the cheapest capable agent models you can call today take opposite approaches. Tencent Hy4, released on November 4, 2026, is Tencent's flagship model: it can read screens and operate software, and Tencent's own benchmark appendix says it wins 27 of 34 tests against DeepSeek. DeepSeek V4.1 Flash, out since September 10 with an October 17 checkpoint, is the price benchmark of the open-weight world: 30 cents per million input tokens at peak, half that off-peak, with MIT-licensed weights you can download today.

Neither model wins outright. DeepSeek is the better pick for coding agents, batch jobs, and anything where cost per token decides. Hy4 is the better pick when the agent has to look at a screen, click through an interface, or work through Chinese-language office documents. One caveat belongs up front: Hy4 is marketed as an open-weight model, but the full release's weights have not been published yet, so today it is API-only. Everything below reads Tencent's numbers as Tencent's own claims, not independent results.

The two contenders at a glance

*The 770B total and 49B active figures come from the August preview's model card. Tencent's November release post does not restate parameter counts, so treat them as carried over from the preview, not confirmed for the final model.

Hy4 is the first model in Tencent's Hy series with vision and computer use. It can read images and charts and operate command-line and graphical interfaces, including professional tools such as Blender and CAD software, which Tencent says received dedicated training for long-horizon tasks. It is available through Tencent Cloud TokenHub and OpenRouter, and inside Tencent products including Yuanbao, WorkBuddy, and CodeBuddy, with a cheaper Hy4 Flash coming soon, according to Tencent.

DeepSeek V4.1 Flash is the efficient tier of DeepSeek's V4 family. Its asymmetric design activates only 8B parameters when reading input and 16B when writing output, which is the structural reason it is so cheap and fast, and its key-value cache uses about 890 bytes per token, roughly a quarter of the previous V4 Flash. The 1017 checkpoint is a minor update with, according to DeepSeek's model card, substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. The API alias is deepseek-flash.

Pricing: Hy4 costs two to three times more

Price is where the two models separate most clearly. At list price, Hy4 charges $0.834 per million input tokens and $2.501 per million output tokens, against DeepSeek's peak $0.30 and $1.20. That is about 2.8 times DeepSeek on input and 2.1 times on output. Cached input is $0.042 for Hy4 against $0.006 for DeepSeek.

Two discounts narrow the gap. Tencent Cloud TokenHub sells Hy4 at 50% off through December 31, 2026, bringing it to $0.417 input and $1.251 output. Discounted, Hy4's output price is only about 4% above DeepSeek's peak output price. DeepSeek, meanwhile, halves its prices off-peak — $0.15 input and $0.60 output outside peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays).

For 100 million input tokens and 20 million output tokens, cache aside: Hy4 at list price costs about $133, Hy4 at the 50% discount about $67, DeepSeek at peak about $54, and DeepSeek off-peak about $27. The discount window is the cheapest time to evaluate Hy4. After December 31, the gap returns to roughly two to three times, and Hy4 only makes financial sense if you need what DeepSeek cannot do.

Benchmarks: read Tencent's 27–7 with care

The headline figure comes from the appendix to Tencent's Hy4 release page, where both models have results for 34 benchmarks. Hy4 wins 27 and DeepSeek wins 7, according to Tencent. DeepSeek's seven wins are concentrated in coding and automation: DeepSWE (77.0 vs. 74.9), SWE-Marathon (46.3 vs. 43.8), Hy-CompanyBench (79.5 vs. 78.8), AutoCath (58.1 vs. 57.2), AutomationBench (60.5 vs. 56.2), PPTOGen (62.5 vs. 60.9), and Foundation Model Bench (36.3 vs. 28.1). Hy4's largest leads are where its new capabilities sit: computer use on OSWorld 2.1 (78.9 vs. 68.4), visual question answering on BuddyVQA (55.6 vs. 43.7), instruction following (79.9 vs. 68.7), E-Bench 2.0 (70.0 vs. 66.8), and APEX-Agents for office tasks (75.2 vs. 72.1).

That scoreline needs three qualifications. First, all 34 results were rerun or compiled by Tencent itself. The DeepSeek column is not DeepSeek's own reporting, and no independent leaderboard has scored the final Hy4 release. Second, Tencent's own publication disagrees with itself: DeepSeek scores 40.0 on Terminal-Bench 4.0 in the main chart and 42.3 in the appendix, and AutomationBench shows 54.3 in one place and 60.5 in another. A lead of one or two points is best treated as noise, and many of Hy4's 27 wins are by about a point. Third, 20 of the 34 tests are Tencent's internal suites. Tencent itself acknowledges that Hy4 "still has a significant gap" against top closed-source models on frontier mathematics, long-horizon development, and visual question answering.

The independent evidence covers only the Hy4 preview, not the final release, and it points the same way: DeepSeek is stronger at coding, Hy4 at math and general tasks. On Vals AI's 19 shared benchmarks, the Hy4 preview won 12 and DeepSeek won 7. On CWE-bench's 250 questions, DeepSeek leads 57.6% to 47.3%, at $0.04 per run against $0.43. On the LMArena leaderboard from November 2, DeepSeek sits 21st and the Hy4 preview 24th. A Tencent-run panel of 213 experts graded 579 real financial tasks: Hy4 3.04, DeepSeek 2.97 on a four-point scale. Tencent also claims Hy4's hallucination rate fell from 15.76% in the preview to 5.45% in the final release; treat that as a vendor claim until an independent test repeats it.

Where each one actually wins

For coding agents and long-running engineering, pick DeepSeek. It leads on DeepSWE and SWE-Marathon in Tencent's own appendix, and the third-party coding tests agree. It also supports up to 384K output tokens against Hy4's 64K, so whole-repository generation and long reports do not need splitting. And it is faster: about 182 tokens per second in reasoning mode and 256 in non-reasoning mode on Artificial Analysis, against roughly 55 for Hy4 on OpenRouter.

For screens, software operation, and office documents, pick Hy4. Its OSWorld lead of 10.5 points is the largest gap in the appendix, and it is the only one of the two that can operate a graphical interface. The E-Bench and APEX-Agents leads point the same way: Hy4 is built for agents that act in software, not just produce text. Its strength in Chinese-language office workflows is a further reason to pick it inside Tencent's ecosystem.

For self-hosting and high-volume batch work, pick DeepSeek. Its weights are available now under the MIT license, it activates far fewer parameters per token, and its smaller key-value cache makes long-context serving cheaper. Hy4's weights are promised with a system card, but until they appear, any self-hosting plan depends on a file that does not exist yet. Batch work that can be scheduled off-peak gets DeepSeek at half price, the cheapest capable inference in this comparison.

One exception: if your office work mainly means generating files rather than operating an interface, DeepSeek's narrow PPTOGen lead (62.5 vs. 60.9) suggests the two are close, and price should decide. Hy4's advantage is in doing things on screen, not in writing documents.

How to try both

The simplest way to compare them is OpenRouter, which hosts both behind one API key: tencent/hy4 and deepseek/deepseek-v4.1-flash-1017. Its Hy4 page lists the 50%-off pricing, a 1,048,576-token context window, and up to 65,536 completion tokens, with tool calling and structured outputs supported.

DeepSeek's own API is the other obvious route, using deepseek-flash; legacy names still work but are served by the 1017 checkpoint. Hy4 is available on Tencent Cloud TokenHub under the model ID hy4, and WorkBuddy and CodeBuddy offer limited free use from November 4 to 15, 2026. Both models speak the OpenAI Chat Completions and Responses protocols and the Anthropic Messages protocol, so switching generally means changing only the base URL and model name. Three details matter when migrating: Hy4's 64K output limit against DeepSeek's 384K, DeepSeek's thinking mode being on by default with reasoning_effort adjustable from 1 to 100 (official scores were measured at 100), and Hy4's slower speed.

Verdict: who should pick which

Choose DeepSeek V4.1 Flash for coding agents, batch or high-concurrency workloads, self-hosting, or cost optimization. It is cheaper at every hour of the day, faster, and its weights are downloadable today under the MIT license.

Choose Tencent Hy4 if your agent needs to see screens and operate software, work through Chinese-language office documents, or answer visual questions. Its computer-use capability is something DeepSeek simply does not offer, and the 50% discount through December 31, 2026 makes this the cheapest window to evaluate it. Do not plan self-hosting around Hy4 until the weights actually appear.

The safest next step is the same for both: run 20 to 50 of your own real tasks through each model and compare cost, speed, and success rate. Tencent's 27–7 is a useful reference, but it is Tencent's own scoreboard, and the independent tests say the coding race is closer than that headline suggests. Your own data beats any leaderboard.

Sources

• Tencent Hy4 release page — hy․tencent․com
• DeepSeek API docs: Models & Pricing — api-docs․deepseek․com
• OpenRouter Hy4 model page — openrouter․ai
• Hy4 vs DeepSeek V4.1 Flash comparison — help․apiyi․com