Hy4 vs DeepSeek V4.1 vs Kimi K3.1: Best Open Model?

By GameDrop Staff · Nov 06, 2026 · Versus · 2 min read

Image: Tencent

Tencent's Hy4 arrived on November 4 with big claims, but which open model should you actually use? Independent testing from Vals AI and Artificial Analysis gives a clearer picture than Tencent's own charts: Kimi K3.1 Pro leads the open-weight pack, DeepSeek V4.1 Flash is the budget pick, and Hy4 sits fourth on the Vals Index.

Vals Index: Kimi K3.1 Pro 62.52%, DeepSeek V4.1 Flash 1017 61.26%, GLM 5.3 59.73%, Hy4 59.35%.
Cheapest per test: DeepSeek V4.1 Flash at $0.657.
Hy4 weights are not public yet; Tencent says they will follow.
Tencent's launch tables mostly use its own test runs, so treat them as vendor claims.

How do Hy4, DeepSeek V4.1 Flash and Kimi K3.1 Pro compare?

Model	Vals Index	Cost per test	List price per 1M tokens (in/out)	Context
Kimi K3.1 Pro	62.52%	$6.44	$2.00 / $10.00	1M
DeepSeek V4.1 Flash 1017	61.26%	$0.657	$0.30 / $1.20	1M
GLM 5.3	59.73%	$8.525	$1.40 / $4.40	1M
Tencent Hy4	59.35%	$1.72	about $0.90 / $2.70	1M

Figures come from the Vals AI index and a roundup of OpenRouter list prices as reported by AI Magazine. Prices change quickly, so check your provider before committing.

Where does Hy4 win?

Tencent says Hy4 ranks first among open-weight models in a blind finance test of 579 tasks judged by 213 professionals, scoring 3.04 against 2.97 for DeepSeek and 2.68 for Kimi. It also says its hallucination rate on one internal test fell to 5.45% from 15.76% on the preview. Vals AI backs up strength in quantitative finance and tax tasks, though it also logged legal-research regressions caused by broken tool calls and timeouts.

Where does Kimi K3.1 Pro pull ahead?

On terminal and agent work. In Tencent's own appendix, Terminal-Bench 4.0 reads Hy4 44.4 versus Kimi K3.1 Pro 66.8 and Claude Opus 5.5 66.4. On the long coding test SWE-Marathon, Kimi scores around 51.7 to 53.8 depending on the run, against 43.8 for Hy4. Kimi also takes vision and computer-use tests like OSWorld 2.1 by a small margin (79.6 versus 78.9). Read our Kimi K3.1 Pro explainer for the model itself.

Tencent's Hy4 benchmark charts — Image: Tencent

Is DeepSeek V4.1 Flash the best value?

For most people, probably. It scores 1.91 points above Hy4 on the Vals Index while costing well under half as much per completed test ($0.657 versus $1.72), and its list price is the lowest of the four. Tencent's own table shows it ahead of Hy4 on SWE-Marathon (46.3 versus 43.8), although scores for the same model differ between Tencent's chart and its appendix.

Which open model should you pick?

Best overall quality and agents: Kimi K3.1 Pro, if you can afford $10 per million output tokens.
Cheapest solid option: DeepSeek V4.1 Flash.
Finance, tax and spreadsheet-style work, with a Tencent ecosystem: Hy4, once the weights land.
Local use: only Kimi and DeepSeek have weights you can download today; Hy4 does not yet.

For the Hy4 launch details, see our Hy4 launch guide, and for closed models see best AI for coding.

What is the verdict?

Benchmarks are a guide, not a guarantee. Hy4 is a clear step up from its own preview but not the open-model leader that launch headlines suggested. Test two models on your own prompts before switching.