Tencent expands its open-weight strategy with Hy4 as it competes on model efficiency
We forecast a 60% chance that official tencent/Hy4-flash rolling 30-day Hugging Face API downloads remain below 68,001 on 21 December 2026.
By 21 November 2026, Tencent's Hy4 flash had 563 likes on Hugging Face, nearly as many as MiniMax M3.1 Flash's 611. Yet Tencent had 5,588 downloads against MiniMax's 150,581. The counts cover roughly four and 35 days, respectively, so MiniMax had a head start. A like resembles stopping at a showroom window, a download resembles asking for the keys, and paid inference is when the invoice arrives. Developers still need evidence from real workloads before treating interest in the weights as evidence of commercial adoption.
Tencent launched hosted Hy4 on 4 November 2026, then released the flagship weights on 20 November. Flash weights had appeared on 17 November. The flagship has 770 billion total parameters, 49 billion active for text and a one-million-token context. Flash has about 220 billion total parameters and 12 billion active, with the same one-million-token context. Distribution Deficit is the gap between available weights, official downloads, real deployments and paid usage; our test counts only official tencent/Hy4-flash rolling 30-day Hugging Face API downloads, not deployments or paid usage.
Tencent's flash announcement lists $0.070 per million input tokens, $0.281 per million output tokens and $0.003 for cached input. TokenHub offers 50% off and WorkBuddy free access through 30 November 2026. Applying that discount to output makes a scenario of $0.1405 per million tokens, saving a buyer the same amount against list price. Tencent's own Flash efficiency and cost figures, including WorkIndex, cost per task and tokens per second, have not been independently tested: it says Flash needs one-sixth of Hy4's KV cache, decodes above 150 tokens per second and costs one-twentieth as much per WorkIndex task. That makes an inexpensive trial a rational decision, particularly for repeated workloads.
Tencent president Martin Lau warned US chip controls could lengthen the next training cycle, a supply-side counterweight to Flash's efficiency-priced adoption thesis. Tencent's September 2026 quarterly filing records RMB81,433 million in capex versus RMB12,983 million a year earlier, or 6.27 times as much and 39.3% of quarterly revenue. Free cash flow was negative RMB69.1 billion, while cloud-services revenue grew 35% year on year. Tencent already has a route to paying cloud customers, but its compute purchases require recurring profitable usage rather than a spike in free trials.
Artificial Analysis scored the flagship at 35 on 3 November 2026, before the weights arrived, against Kimi K3.1 Pro's 46 and Qwen4 Max's 43. It described Hy4 as "somewhat expensive when comparing to other open weight models of similar size". The evaluation recorded 160 million output tokens and 14.1 minutes per task. GPT-6.2 Sol scored 57 at $1.01 per task, compared with Hy4's $1.03 at 35. Tencent's own flagship card admits "a notable gap remains between it and top proprietary models". Open weights expand who can run Hy4; the existing score still shapes what customers will pay.
The strongest case for a faster adoption curve is the flash product itself. Tencent's flash card puts WorkIndex at 42.1, above Hy4 preview's 39.7 and below full Hy4's 51.4. In an internal blind exercise, 85 users assessed 259 tasks, awarding flash satisfaction of 2.79 out of four versus Claude Haiku 5.5's 2.81, at RMB0.26 per task. Tencent also claims one-eighth of DeepSeek V4.1 Flash's cost per task, while saying flagship Hy4 leads open-source models on Agent Arena. Those figures give cost-conscious hosts a reason to fetch smaller weights.
Tencent's promotions give early curiosity somewhere to go. Flash's 563 likes approached MiniMax M3.1 Flash's 611 in a fraction of the time. A 50% TokenHub discount and free WorkBuddy access give developers a reason to experiment now; Tencent says OpenRouter integration is coming soon. Downloads of the original repository by direct users, third-party hosts and teams preparing quantisations would all count. Those routes can accelerate in the remaining weeks without waiting for a new model release.
The peer record makes that case formidable. All seven Chinese open-weight repositories in the comparison registered at least 68,001 rolling 30-day downloads on 21 November 2026. Kimi K3.1 Pro reached 68,001 at 19 days old; DeepSeek V4.1 Flash reached 511,124 at 35 days; MiniMax M3.1 Flash reached 150,581 at 35 days. Older MiniMax M3, Kimi K3 and Qwen3.8 Flash Next recorded 147,130, 731,025 and 2,601,320. Tencent has to compete for a distribution habit other labs already command.
And Z.ai's GLM-5.3 provides the clearest example of a flash model leading its flagship. At 88 days old, GLM-5.3 had 1,462,339 rolling 30-day downloads; its flash tier had 6,400,766, or 4.4 times as many. The mechanism is appealing: an accessible version can reach more deployments than the large original. These older repositories have had months to build distribution, while Tencent is approaching its first month. That maturity makes Tencent's opening pull rate the more useful figure for a deployment decision.
Extending Tencent's opening 5,588 downloads at a constant pace implies about 42,000 over a month, so hosts should not mistake initial interest for established distribution. Kimi reached that mark in 19 days, which makes the opposing case harder to dismiss. Hy4 flagship alone had 974 downloads on its first day; flash's initial slope still calls for acceleration. That acceleration can arrive through original weight pulls by developers, hosts or quantisers, even if their later derivative downloads and hosted calls fall elsewhere. Distribution Deficit can narrow through downloads while Tencent's cash returns still depend on billable work.
Licensing adds another complication to the apparent giveaway. Under Section 10 of Tencent's Hy licence, model-as-a-service operators need separate written permission above $50 million of worldwide gross revenue in 12 months or 100 million monthly active users; internal use is exempt. A host that attracts enough customers can end up negotiating with its model supplier. At the 21 November snapshot, flash FP8 had 1,535 downloads and base another 826, neither counted in the official flash total. Tencent can gain reach beyond its own API while leaving cash capture to later contracts and serving decisions.
WorkBuddy free access ends on 30 November 2026, when the official flash download count becomes more telling than likes. A fourfold rise from 5,588 to 22,352 official Hy4-flash rolling 30-day downloads would give prospective hosts stronger evidence of developer uptake. That signal would carry more weight than likes, while cloud growth of 35% offers Tencent an existing commercial outlet for managed inference. Flash's $0.070 input and $0.281 output per million tokens make official tencent/Hy4-flash rolling 30-day Hugging Face API downloads a proxy for developer adoption of an efficiency-priced open model. Artificial Analysis puts Hy4 at $1.03 versus Claude Fable 5.2's $26.54 per Intelligence Index task (scores 35 and 65); downloads cannot establish Flash quality independently because no independent Flash score exists. Providers charging that premium face a harder negotiation wherever the cheaper model is sufficient.
We forecast a 60% chance that official tencent/Hy4-flash rolling 30-day Hugging Face API downloads remain below 68,001 on 21 December 2026; buyers should test the advertised price while investors demand evidence of paid conversion. We would be wrong if that same official repository's rolling 30-day Hugging Face API count reaches at least 68,001 on 21 December 2026 through eligible developer, host or quantiser pulls.