What powers the open frontier

figure 1 · the lineage

Same recipe, new proportions

Put the arc’s pipeline next to a 2026 frontier run and the skeleton is unchanged: curate, imitate, then optimize against feedback. What changed is the proportions. In InstructGPT, RL was a short final polish; in GLM-5 the post-training pipeline is three RL stages deep — reasoning, agentic, general — followed by a distillation pass to knit the stages together. And the feedback itself has migrated: away from learned reward models toward verifiable rewards (entry 18’s RLVR) wherever a checker can be built.

click any figure to zoom · scroll to magnify · drag to pan

figure 2 · GLM-5’s assembly line

SFT, then RL three times, then distill it back together

GLM-5 (744B parameters, 40B active per token, trained on ~28.5T tokens) runs post-training as a sequence: SFT — with interleaved and preserved “thinking” modes, and error segments masked in agent trajectories so the model learns to recover from mistakes rather than imitate them. Reasoning RL — GRPO with binary outcome rewards on math, science, code. Agentic RL — long-horizon tasks in real execution environments. General RL — style, factuality, safety, with hybrid rewards. Finally, on-policy cross-stage distillation: the specialist checkpoints teach one final student, so the reasoning gains and the chat gains stop being a trade-off.

The difficulty filtering in reasoning RL is worth pausing on: training problems are kept only if the previous model generation rarely solves them but stronger teacher models can — the curriculum is pinned, permanently, at the edge of current ability. That is entry 14’s curation instinct, automated.

figure 3 · the RL factory

Generation and training, finally divorced

The infrastructure story is as important as the algorithm. Long agentic rollouts made synchronous RL (generate a batch, wait, update, repeat) untenable — one slow trajectory stalls the fleet. GLM-5’s answer (the open-sourced slime framework) is a fully asynchronous factory: inference engines generate trajectories continuously on their own GPUs; training engines consume them as they arrive; weights sync every few updates. Correctness patches — exact token round-tripping, double-sided importance sampling, dropping trajectories from too-stale model versions — keep learning stable while the two halves drift.

figure 4 · the reward stack

Ten thousand environments and a hybrid judge

Where entry 15 had one reward model and entry 18 had three verifiable domains, GLM-5’s agentic stage trains against 10,000+ verifiable scenarios — software engineering across nine languages, Docker-synthesized terminal tasks, multi-hop search — where the reward is “did the tests pass, did the task complete.” For the subjective remainder, a hybrid stack: rule checks where possible, outcome reward models where not, generative judges above that — and expert-written responses kept in the loop as stylistic anchors so the model doesn’t converge on judge-pleasing sludge.

the through-line every layer of that stack is a hedge against reward hacking — entry 15’s oldest lesson, now an architecture. The verifiable floor keeps growing precisely because verified reward is the one signal that cannot be flattered.

figure 5 · K3’s bet

2.5× the intelligence per unit of compute

Kimi K3 attacks a different margin. It is the largest open-weights model ever released — 2.8T parameters — but its headline is that only 104B are active per token, and every architectural choice buys intelligence per FLOP: Kimi Delta Attention interleaved with full attention 3:1 (cutting KV-cache ~75% on the way to a native 1M context), attention residuals for cross-layer information flow, and a LatentMoE that stays stable while activating just 16 of 896 experts. Moonshot’s claim: ~2.5× the scaling efficiency of Kimi K2 — the same intelligence for 40% of the compute, or more intelligence for the same bill.

Post-training rhymes with GLM-5’s: large-scale task synthesis across reasoning, general agents and coding agents, RL infrastructure built for million-token contexts, and preserved thinking promoted to an API contract — the reasoning chain must round-trip through the caller or quality degrades. Alongside the weights, Moonshot opened the machinery: FlashKDA kernels, the MoonEP communication library, and AgentEnv, the sandbox fleet the agents trained in.

figure 6 · worked by hand

The sparsity arithmetic

The numbers that define the era, divided out: K3 activates 104B of 2,800B parameters — 3.7%. GLM-5 activates 40B of 744B — 5.4%. Compare the dense 175B GPT-3 of entry 15: 100% active, every token. And the context ladder these models train up — 4K to 200K to 1M — is its own quiet scaling law, bought stage by stage in mid-training.

results · the standings

Open weights, closed-model scores

The scoreboards say the gap has effectively closed on many axes. GLM-5 is the first open-weights model to reach an Artificial Analysis Intelligence Index of 50, and sits atop the open-model arena boards; its successor GLM-5.2 ships under MIT with a usable 1M-token window. K3’s report tables it against the strongest closed models of July 2026 — ahead on some (GPQA Diamond 93.5, BrowseComp 91.2), behind on others (agentic coding suites) — which is itself the story: the comparison is now per-benchmark, not per-league.

One asymmetry to keep: these frontier reports open the recipe and the infrastructure, but — unlike Tulu 3 — not the training data. The last mile of transparency is still where entry 18 left it.

the equations · worked

Compute per token, when most of the model sleeps

applied · our reading

Reading the frontier from a small team’s desk

• The recipe converged; the moat moved. Both giants run the arc you just read — curate, SFT, preference/RL, verify — at scale. What separates them is environments, reward engineering, and RL infrastructure. For a domain team the translation is direct: your moat is your verifiers and your expert data, not your training loop.
• Build checkers, get frontier-style training free. GLM-5’s 10k environments are, structurally, unit tests with a reward attached. A regulated domain (returns that must reconcile, filings with hard formats, statutory deadlines) is rich in checkables — the RLVR pattern from entry 18 scales down to one GPU and up to 744B unchanged.
• Efficiency per FLOP now decides sovereignty. K3’s 3.7% activation and GLM-5.2’s MIT license mean frontier-class capability is, for the first time, something you can legally and physically host. The entry 17 question — “what fits on hardware we control?” — is now being answered from the model side too.

a dating note this entry describes the frontier as of July 2026 and will age faster than entries 14–18 — that is what a coda is for. The pipeline lessons are durable; the scoreboard is not.