Hy4 vs Open-Weight Rivals: How to Choose Between Kimi, DeepSeek, GLM and Inkling

Hy4's launch timing was delicate: on the same day (Nov 4), Kimi K3.1 Pro also published its weights, and DeepSeek V4.1 Flash and GLM-5.3 had arrived in the preceding weeks. The open-camp flagship race suddenly became a four-way contest in early November. This article combines Tencent's official chart with Artificial Analysis' independent index to take the contest apart — conclusion first: nobody wins everywhere, and scenario-based selection is the right approach.

The 6–6 in the official chart: Hy4 vs Kimi

In Tencent's official 12-item chart, Hy4 and Kimi K3.1 Pro tie 6–6 (apiyi's rank-by-rank count). Hy4 leads in code generation, 3D modelling, retrieval and research tasks; Kimi leads in terminal operation, automation and science-terminal tasks. The widest gaps: Terminal-Bench 4.0, Kimi 66.8 vs Hy4 44.4; AutomationBench, Kimi 65.6 vs Hy4 56.2; Agents' Last Exam, Kimi tied with Claude at 43.4, Hy4 41.0. On OSWorld 2.1, Kimi's 79.6 is also slightly above Hy4's 78.9.

The independent index widens the gap: AA index Kimi 46, Hy4 35, an 11-point difference; AA-Omniscience hallucination rate Kimi 37%, Hy4 89%. But Hy4 has its own wins: FrontierCode 2.0 (AA) Hy4 31%, best among Chinese labs, Kimi 25%; the finance crowdsourced test Hy4 3.04, Kimi 2.68. And one key difference: Kimi's weights shipped on Nov 4 itself; Hy4's still haven't.

DeepSeek V4.1 Flash: cheap, but also weak on terminal tasks

AA index DeepSeek V4.1 Flash 32.6, slightly below Hy4's 35; AA Terminal-Bench 4.0 33.8% (Hy4 37%), Tencent's table 40.0% (Hy4 44.4%) — Hy4 leads by about 4 points under both harnesses. In the finance crowdsourced test DeepSeek scores 2.97, just 0.07 behind Hy4's 3.04, the closest rival. DeepSeek's strengths are its ecosystem, its reputation for value and its already-open weights; its weakness is terminal tasks, where it lags Kimi and the closed flagships just like Hy4.

GLM-5.3: actually ahead of Hy4 on terminal tasks

GLM-5.3 scores 41.9% on AA's Terminal-Bench 4.0, above Hy4's 37%; in Tencent's table 42.9%, slightly below Hy4's 44.4% — the two harnesses give opposite verdicts, which means the two models are roughly in the same tier on terminal tasks and the harness decides the winner. In the finance crowdsourced test GLM scores 2.68, tied with Kimi, behind Hy4's 3.04. GLM-5.3-Flash (AA 27.8) is the lightweight tier, closer in positioning to Hy4 Flash. GLM's weights are already open — a clear advantage over Hy4.

Inkling 2 Small: a different race for a smaller model

Thinking Machines Lab's Inkling 2 Small (594B total, 17B active) scores AA index 31 and Terminal-Bench 4.0 of just 13.1% — not in Hy4's tier. But its API price is $0.30 / $1.00, nearly two-thirds cheaper than Hy4's list price, and its weights are open under Apache 2.0, with the NVFP4 checkpoint on Hugging Face. Inkling is positioned as "good enough and cheap"; Hy4 as "the all-round open-camp flagship" — budget-first buyers pick Inkling, capability-first buyers pick Hy4.

Scenario-based selection

The gap to closed flagships: not at the same table yet

In the official chart, Hy4's gap to Claude Opus 5.5 and GPT-6.1 Sol is widest on the hardest tests: Hy-FrontierMath T5 13.0 vs GPT's 97.5, Goal2Repo 23.2 vs Claude's 74.0, Terminal-Bench-Science 19.8 vs GPT's 60.9. On the AA index, Hy4's 35 ties Grok 4.7 (xhigh), but the closed flagships' top scores are above 50. Tencent admits the gap on its own launch page — competing inside the open camp is one thing; catching the closed models is another.