GLM-6.0 Full Self-Training Roadmap Deep Dive: Environment Relay, Stopping Criteria, and a Comparison of Three Recursive Carriers

> Anchoring event: Zhipu's 2026-08-31 interim earnings call, where Tang Jie defined GLM-6.0 as Full Self-Training (= Recursive Self-Improvement). Supporting sources: Tencent Hy4 preview announcement (2026-08-28), Hunyuan Hyra-1.0 self-improving agent (2026-07-21), THUDM/slime (Apache-2.0, 2026-06), GLM-5.3 third-party analysis, Sakana Darwin Gödel Machine paper (arXiv:2505.21444).

Confidence labels used below: CONFIRMED (official leaderboards/papers/downloadable artifacts) ｜ VENDOR (vendor-reported, not independently reproduced) ｜ RUMOR (customer testimony/third-party relay).

Key points at a glance

• 743B: identical base parameters for GLM-5.3 and GLM-5.2, ~40B active (MoE)
• +50%: claimed end-to-end completion gain on Z.ai Code Bench via post-training only, same base
• +20pp: DeepSWE v1.1, <45% → 66.9%, same base, expanded environments only
• 2.99: Tencent Hy4 preview internal blind score (/4.00); GLM-5.3 = 2.92, Kimi K3 = 2.94
• +31.8%: throughput gain from Hy4's self-optimizing inference stack
• 20→50%: Sakana DGM on SWE-bench — the empirical ceiling for code-modifying RSI so far

1. What the debate is about

In late August 2026, China's second-tier open-source labs shifted from competing on benchmarks to competing on "making themselves improve benchmarks" — each with a different recursive carrier:

• Zhipu bets on the environment: one GLM generation builds the training ground; the next trains inside it.
• Tencent Hy4 preview bets on the self-stack: the model optimizes training methods, data strategies, evaluation, and low-level operators.
• Sakana (Darwin Gödel Machine) bets on code itself: the model edits its own Python source.
• DeepMind AlphaEvolve bets on narrow-domain evolution: evolutionary search over specific algorithmic functions.

2. Fact sheet

| Item | Content | |---|---| | Trigger | Tang Jie, 2026-08-31: "GLM-6.0 = Full Self-Training" | | Technical positioning | "Scaling deep while scaling up" — post-training scale, long-horizon task environments, and task feedback expanded together | | Zhipu's recursive carrier | Training environments (task distribution + verifier + data generation functions), elevated from consumable to pipeline-standard product | | Empirical baseline | GLM-5.3: same 743B base + same MoE architecture, post-training only → Z.ai Code Bench +50%, DeepSWE v1.1 <45% → 66.9% | | Key infrastructure | THUDM/slime (Apache-2.0) — Megatron-LM + SGLang + Data Buffer, 11+ downstream projects | | Tencent comparison | Hy4 preview (2026-08-28): 770B / 49B active / 1M context; first disclosed RSI loop, +31.8% end-to-end inference throughput | | International comparison | Sakana DGM: SWE-bench 20% → 50%; AlphaEvolve: 48 scalar multiplications for 4×4 complex matrices (breaking Strassen 1969) | | Corrective stance | Tang Jie: "parameter scale is not the only variable in Scaling" | | Risks | Verifier drift / "boxed 1 collapse" / RL signal self-reinforcement — knowing when to stop and how to correct is the hardest part | | Evaluation window | VENDOR-labeled results await independent reproduction, expected 4–6 weeks after GLM-5.3 weight release (~2026-08-28) |

3. GLM-5.3: what "+50% on the same base" means

Vendor-reported (VENDOR) results released 2026-08-14:

| Benchmark | GLM-5.2 | GLM-5.3 | Change | Measures | |---|---|---|---|---| | Z.ai Code Bench (Max effort) | baseline | +50% | — | internal end-to-end coding completion | | DeepSWE v1.1 | <45% | 66.9% | +>20pp | real software engineering | | Terminal-Bench 3.0 | 4.6% | 28.3% | +23.7pp | long-horizon CLI / agentic coding | | CyberGym | 77.2% | 84.5% | +7.3pp | automated vulnerability discovery | | ExploitBench | 24.4% | 54.4% | more than doubled | exploit chains | | HLE (with tools) | n/a | 62.5% | — | expert reasoning + tool use | | GDPval-AA v2 (Elo) | n/a | 1,769 | — | economically valued tasks | | Agents' Last Exam | n/a | 28.5% | — | hard, broad agentic reasoning |

Caveats the author flags:

• Z.ai Code Bench is an internal benchmark — the +50% is relative (GLM-5.3 = 33.7% vs GLM-5.2 = 23.4%), not a pass rate jump from zero.
• On DeepSWE, GLM-5.3 (66.9%) still trails Kimi K3 (69%) and DeepSeek V4-Pro (80.6% SWE-bench Verified) — second tier among open models.
• Terminal-Bench 3.0 and 4.0 use different task sets; the 4.6→28.3 jump is not cross-version comparable.
• The ExploitBench doubling raised dual-use concerns, reportedly triggering a temporary safety-review pause on weight release.

Engineering significance: if expanding environments alone yields large gains on an unchanged base, then letting the model expand its own environments (GLM-6.0's bet) pushes the same lever one notch further.

4. Deconstructing Tang Jie's three claims

Definition: GLM-6.0 will feature self-purification across pre-training / mid-training / post-training, autonomously judging "when to stop training" and correcting its own errors — Full Self-Training, an industrial version of RSI.

1. Environment self-construction: "environments go from consumables to pipeline-standard products"; "weights can be open-sourced, the environment synthesis pipeline cannot be copied." GLM-5.3 already powers an Infra Agent for inference-kernel and service-stack optimization; GLM-6.0 would hand the environment production line to the model itself. 2. "Scaling deep while scaling up": parameter scale is not the only variable; "simply making parameters bigger is not the essence of Scaling." Resources currently tilt toward post-training and long-horizon RL rather than a 3T base. 3. Engineers are still in the loop: the interim report admits the pipeline retains substantial manual steps — a more open admission than Tencent's Hy4, which claims an initial closed loop (model proposes → experiment → adjust code/logs → feedback) with a quantified +31.8% throughput gain.

5. Four recursive carriers compared

| Lab | Recursive carrier | What it modifies | Evidence | Closed loop? | |---|---|---|---|---| | Zhipu (GLM-5.3→6.0) | Environment | task distribution + verifier + data generation | GLM-5.3 +50% on same base | semi-closed, engineers in loop | | Tencent Hy4 preview | Self-stack | training methods, data, evaluation, operators | +31.8% inference throughput | initial closed loop | | Sakana DGM (2025-05) | Code | its own Python source | SWE-bench 20%→50%; Polyglot 14.2%→30.7% | research prototype, human gatekeeping | | DeepMind AlphaEvolve (2025-05) | Narrow-domain algorithms | specific math/engineering functions | 4×4 complex matrix: 48 multiplications; +23% Gemini training kernels | research-grade, auto evaluation | | CMU SRT (2025-05) | Reward signal | self-consistency as intrinsic reward | Llama-3.1-8B MATH-500 52.6%→~60% | research-grade; reward hacking observed | | OpenRSI / Frontis-MA1 (2026-08-20) | Evolutionary search | programs, not systems | MLE-Bench Lite 39.4%→71.2% | meta-evolution; authors say not yet RSI |

The author's takeaway: modifying environments offers high ceilings and partial open-sourcing but requires industrial pipeline capability; modifying the self-stack closes loops quickly; modifying code is purest but blocked by safety/alignment; narrow-domain search is most practical but limited to verifiable functions.

6. Correcting the "RSI needs 3T parameters" claim

Of three popular claims evaluated, the first (RSI's hardest problem is anti-cheating and overfitting, requiring objective anchors) is confirmed by the literature: SRT's "boxed 1 collapse" shows RL signals drifting from correctness-correlated to confidence-correlated; DGM documents reward hacking (hallucinated tool calls, removed detection markers); AlphaEvolve itself names the need for automatic evaluators as its biggest limitation.

The claim that RSI requires ~3T parameters is the weakest: small/medium models (Qwen 2.5/3, Llama 3.1-8B) have demonstrably self-trained; reward-signal drift is a verifier/reward-engineering problem, not a parameter problem; and Zhipu's own messaging explicitly rejects "simply making parameters bigger" as the path.

> Corrected thesis: RSI feasibility shows no strong evidence of depending on parameter count; the real bottleneck is verifier robustness and letting the model judge when to stop. Zhipu's GLM-6.0 bet is not scaling to 3T but industrializing the "environment relay."

7. The slime framework: engineering foundation

slime (Apache-2.0) splits RL post-training into three modules:

• Training — Megatron-LM (TP+PP+SP)
• Rollout — SGLang Engine + Router (high throughput, KV-cache reuse)
• Data Buffer — the sole communication medium, via Ray

Engineering facts: reportedly trained GLM-5.2 in ~2 days (VENDOR); supports 744B+ MoE; implements PPO / GRPO / REINFORCE++ / GSPO / OPD; 11+ downstream projects (Dressage, Miles, vime, Relax, OpenClaw-RL, P1, TritonForge, APRIL, qqr, ART, RLVE). Design philosophy: deliberately no multi-backend abstraction — deep SGLang-only optimization. The core loop stays fixed; each GLM generation swaps in a new data-generation function.

8. Timeline assessment

Based on Zhipu's historical cadence (~60–80 days per minor release; ~30 days for same-base "polish" releases) and Tang Jie's statement that direction is set but timing is not:

| Window | Preconditions | Likelihood | |---|---|---| | 2027 Q1 (fastest) | environment pipeline proven by 2026 Q4; GLM-5.4/5.5 as polish releases | low (~15%) | | 2027 H2 (conservative) | "environment self-construction is our next step" + engineers still in loop | most plausible per author |

The author stresses that any specific version-count forecast is commentary, not an official Zhipu roadmap, and that all VENDOR-labeled numbers await independent reproduction.

*Source: zhichai.net forum post (translated and summarized). Claims flagged VENDOR/RUMOR above have not been independently verified.*