Hy4 Is Real. ‘Recursive Self-Improvement’ Isn’t What Tencent Did.

On August 28, Tencent released Hy4 preview: a 770-billion-parameter mixture-of-experts model, open-sourced under Apache 2.0, with a context window exceeding one million tokens and benchmarks that land it at seventh place out of 228 models on the BenchLM aggregate leaderboard. GPQA Diamond at 92.3. SWE-bench Pro at 65.7%. The model is real, the weights are genuinely available, and the Apache 2.0 license means there’s no revenue threshold before commercial use becomes a conversation with a Tencent lawyer. The open-source AI ecosystem needed another serious entrant at the frontier. It got one. Then the announcement said the model was built using “recursive self-improvement,” and that’s where the story stops being straightforward.

What Tencent Built

The Hy4-preview model card discloses the architecture clearly enough to evaluate. 78 layers, of which 77 are mixture-of-experts blocks each containing 256 routed experts plus one shared expert. Per token, eight routed experts activate alongside the shared one — nine of 257 possible, for a 6.4% activation ratio at 49 billion active parameters out of 770 billion total. That ratio is aggressive by MoE standards — comparable to what DeepSeek V3 achieved with its own sparse architecture — and the practical outcome is a model with the capacity of something much larger, running at the inference cost of something much smaller.

The attention mechanism is explicitly Gated DeepSeek Sparse Attention with IndexCache — a direct borrow from DeepSeek’s architecture, which uses cross-layer sparse index reuse to cut computational overhead without degrading quality. Tencent is not the first to build on this; GLM-5.3 did the same. What Hy4 adds that the others don’t have is Identity Hyper-Connections: four parallel residual streams running alongside the standard residual path, designed to improve information flow across the model’s 78 layers. That’s the architecturally novel element. There’s also a dedicated 10-billion-parameter multi-token prediction layer — 700 million of those active — integrated into serving recipes by default to handle speculative decoding.

The result is a model that the Tencent announcement positioned against Kimi K3 and GLM-5.3 in a blind evaluation by 163 internal experts across 203 engineering tasks. Hy4 averaged 2.99 out of 4.00; Kimi K3 hit 2.94; GLM-5.3 hit 2.92. In direct head-to-head comparison with GLM-5.3, Hy4 won 46.8% of tasks, tied 12.8%, and lost 40.4%. Those are real margins — Hy4 is slightly ahead, not dominant. It is a competitive frontier-class open-weight model, which in August 2026 is not a trivial claim to make.

What “Recursive Self-Improvement” Actually Means Here

The phrase “recursive self-improvement” does specific work in AI safety research. It names the scenario in which an AI system modifies its own training process in ways that produce a better AI, which then further improves the training process, generating a compounding feedback loop. It is the mechanism underlying most serious forecasts of discontinuous AI capability growth. Researchers who study it use the phrase carefully because what it implies — an AI that autonomously upgrades itself past human ability to intervene — is categorically different from an AI that performs well on benchmarks.

What Tencent describes under that label is different. The announcement says Hy4 achieved this by “autonomously analyzing bottlenecks in its inference system and carrying out multiple rounds of optimization on areas such as operator fusion and communication optimization,” resulting in a 31.8% throughput improvement over baseline. The subsequent technical analysis is more precise: “This references improvements to the build process itself — not model autonomy or self-modification during inference. The 31.8% throughput improvement derives from engineering optimization work applied during training infrastructure development, not from the model learning to improve itself.”

In plain terms: during development, Tencent used the model to identify inefficiencies in its own training infrastructure — kernel fusion opportunities, communication bottlenecks in the multi-GPU setup — and implemented the resulting optimizations. That’s a meaningful use of a capable model in an MLOps context. It is also something that has been done at AI labs for years under names like “automated hyperparameter search” and “neural architecture search.” The 31.8% throughput gain is a legitimate engineering achievement. It does not involve the model modifying its own weights, does not create a feedback loop that yields progressively more capable iterations, and does not demonstrate anything that the AI safety community would recognize as recursive self-improvement. The decision to call it that was a choice.

The choice is not without consequence. The engineers who most carefully track the difference between “impressive model” and “self-improving system” are the same ones who will run Hy4 in their pipelines or study its architecture. Using their vocabulary to describe a standard pipeline optimization loop does not make the marketing land better with that audience. It makes the genuine engineering achievement harder to evaluate — because now the first question is how much of the claim is real, rather than what the 31.8% number actually means in production context.

The Benchmark Situation

Hy4’s vendor-reported benchmark numbers are impressive on their face. GPQA Diamond at 92.3 is genuinely high — but BenchLM’s independent aggregate puts Sakana Fugu-Ultra at 95.5, a 3.2-point gap that matters at the frontier. SWE-bench Pro at 65.7% sounds competitive until you note that Claude Mythos 5 sits at 80.3%, a 14.6-point gap that is not recoverable by price advantage. DeepSWE at 64.3 runs 8.4 points behind GPT-5.6 Sol. Humanity’s Last Exam with tools at 55.4 trails Claude Opus 5 by roughly 9 points.

The most significant gap, for anyone building agentic systems, is APEX-Agents at 37.1. That benchmark measures multi-step planning in tool-calling scenarios — exactly the use case for which a 1-million-token context window would matter most. A 37.1 on APEX-Agents from a model with 1M context is a signal worth taking seriously before routing agent workflows through Hy4. The same BenchLM profile does show Hy4 at 83.9% on WideResearch and 83.7% on MCP Atlas — its strongest independently-measured results — which confirms that the model has real strengths in knowledge retrieval and structured tool calling. The profile is uneven, not uniformly weak. APEX-Agents is the specific gap that matters for agentic builders.

What BenchLM’s profile also reveals is that only 28 of 408 tracked benchmark slots have displayable evidence. The model landed four days ago; independent verification takes time. The vendor-reported GPQA Diamond and SWE-bench numbers may hold up under independent testing or may not. The honest characterization right now is that BenchLM ranks Hy4 seventh of 228 models on the evidence that exists, with most potential comparison points still unverified. That’s a strong opening position. It’s also an incomplete one.

The throughput situation is equally honest in what it says about where Hy4 actually fits. At 26 tokens per second with a P50 latency of 3.77 seconds, Hy4 is competitive on price — $2.501 per million output tokens against GPT-5.6 Sol’s $30.00 — but not on speed. For batch workloads, document processing, or offline research tasks, that’s a workable tradeoff. For interactive agent loops where a user is waiting for each step, 3.77 seconds first-token latency is a problem. The technical analysis notes that the model’s output is also capped at 64,000 tokens on OpenRouter despite the 1-million-token context window claim — the model can read a million tokens, but it cannot produce at that scale in the current serving deployment.

None of this makes Hy4 a bad model. It makes it a model with a specific profile: excellent knowledge recall, strong on document-level coding tasks, competitive pricing for high-input/low-output workloads, slower than the frontier on agentic planning. That profile is genuinely useful for a subset of production deployments, and the 52.8 billion yuan quarterly investment Tencent has committed to AI infrastructure provides some confidence that the model will be maintained and improved. The question is whether the next version arrives with standard benchmarks or another round of internal expert evaluation.

Why It Matters That the Language Is Wrong

Hy4’s real position in the open-weight landscape is valuable and clear: a DeepSeek-derived MoE architecture with a genuinely novel residual structure, Apache 2.0 licensing that makes it one of the most permissively licensed frontier models available, and benchmark scores that put it in the top 5% of everything publicly available. That case stands on its own. It doesn’t need borrowed vocabulary from AI existential risk research to be interesting.

The “recursive self-improvement” framing does something specific to how Hy4 gets evaluated. It sets an expectation — of autonomous capability growth, of a system that improves itself past human-directed iteration — that the model does not meet and was never designed to meet. Engineers who encounter that phrase and know what it means will discount the claim immediately and then wonder what else in the announcement is similarly imprecise. Engineers who don’t know what it means may expect something that isn’t there. Neither audience is served.

The actual achievement at Tencent was using a capable model to automate a portion of the MLOps work that normally requires human engineers to profile, identify, and implement infrastructure improvements — and getting 31.8% throughput from it. That’s a story worth telling in its own terms. A 770-billion-parameter open-weight model that helps optimize its own serving infrastructure is already impressive. Calling it recursive self-improvement doesn’t add anything except noise at the frequency where the most important conversations about AI are happening.

AI-generated editorial illustration · TemperatureZero · August 30, 2026