News for hackers. Stuff that matters.
Re:GLM5 Released on Z.ai Platform by jfaganel99 on Thursday February 12, 2026 @03:55AM (#46986406)
The thing nobody seems to be talking about here is "slime," the async RL training framework they built and open sourced (github.com/THUDM/slime). It's been behind GLM-4.5, 4.6 and 4.7 too.

Everyone's debating benchmarks but honestly the actual gap between frontier and non-frontier models right now is RL infrastructure, not pre-training compute. We've known since DeepSeek-R1 that RL on top of a strong base model is where the magic happens, but doing it efficiently at scale is a genuinely hard systems problem. Rollout generation alone eats 90%+ of RL training time, and their APRIL strategy for handling the long-tail problem is a real contribution.

The fact that both the model AND the RL infra are MIT licensed matters way more to me than whether it edges out Opus 4.5 on SWE-bench. Benchmarks are temporary. Training infrastructure compounds.

Re:GLM5 Released on Z.ai Platform by yuvadam on Thursday February 12, 2026 @07:12AM (#46987812)
Do you mind explaining what this RL infrastructure actually entails and why it's a hard problem?
Re:GLM5 Released on Z.ai Platform by pama on Thursday February 12, 2026 @08:02AM (#46988305)
Not OP. Personal opinion on why it is a somewhat hard problem. The main problem is using the available compute correctly and productively while doing two very separate types of tasks that were previously solved independently: generating responses with llm inference engines and modifying weights with a training code. A step of training updates the weights so the inference engines have to adjust theirs, but we talk about 750B parameters and multiple inference servers. Stale weights can be used instead, but only for a tiny bit and the data from them needs special corrections that also involve large compute/memory. Your inference engines better be deterministic (for given pseudoRNG; it clashes with parallelism) or you have a way to correct the probability streams. Ideally inference and training should have same everything at the bit level when they handle the same context, but we dont live in that world yet. And of course, GPUs break. For no great reason, other than the tiny scale of their features making them fragile. And because you scale, you need to handle failures gracefully and efficiently.
Re:GLM5 Released on Z.ai Platform by zozbot234 on Thursday February 12, 2026 @08:20AM (#46988516)
Surely you could just pre-generate rollouts with slightly stale weights and then cheaply verify the rollout when up-to-date weights stream in by treating the former solution as speculative decoding. Sounds quite trivial to me, perhaps I'm missing something.
Re:GLM5 Released on Z.ai Platform by pama on Thursday February 12, 2026 @08:33AM (#46988628)
Cheap verifying of speculative decoding only works for a few tokens at a time. Long sequence generations (thousands to tens of thousands of tokens in typical rollouts for thinking models) are dominated by distribution drift on stale weights (because slightly wrong probabilities multiply over long streams), and the off policy RL training methods dont work well (high variance) for such high dimensional problems.
Re:GLM5 Released on Z.ai Platform by jfaganel99 on Thursday February 12, 2026 @07:52AM (#46988190)
My read: RL post-training loops between generating rollouts (sample responses, score them) and doing updates. The pain is rollouts, since it’s autoregressive token generation and lengths have a long tail, so synchronous setups end up waiting on stragglers while GPUs sit idle.

slime seems aimed at that systems bottleneck: decouple rollout workers (SGLang) from the trainer (Megatron) with a buffer so training keeps consuming while rollouts stream in. APRIL then mitigates the long tail by over-provisioning, taking the first N completions, and carrying partial rollouts forward.

If I’m missing something, curious what others saw.

Re:GLM5 Released on Z.ai Platform by HarHarVeryFunny on Thursday February 12, 2026 @09:53AM (#46989575)
Efficiency of RL training is a cost issue for the organization training the model, but the capability/utility of the trained model depends on the specifics of the RL training data and training regime (as well as the strength of the base model), which seems to be what differentiates the SOTA coding models.

It's neat that Z.ai are opensourcing slime, and are themsleves using DeepSeek's Sparse Attention - a different approach to that of the big US companies.

Re:GLM5 Released on Z.ai Platform by zozbot234 on Thursday February 12, 2026 @10:02AM (#46989710)
Rumor has it that the big proprietary companies are also using DSA now.