Z.ai's GLM-5.3 Beats GPT-5.6 on Coding Without Touching Its Base Weights

Together AI has flagged that GLM-5.3 is now topping GPT-5.6 Sol and Claude Fable 5 on agentic coding benchmarks, with the smaller GLM-5.3 Flash sitting right behind it. The interesting part sits underneath the leaderboard position. GLM-5.3 is a post-training upgrade rather than a newly pretrained base generation, with Z.ai expanding reinforcement learning across more environments and more varied tasks on top of the exact same GLM-5.2 weights.

Same base, different model

The base is identical to GLM-5.2 (744B total, 40B active MoE, IndexShare architecture), and Z.ai states plainly that every gain comes from post-training. No new pretraining run, no parameter-count reveal, and a set of very large benchmark jumps attributed entirely to post-training scale.

Nathan Lambert summarized the Z.ai blog post bluntly: scaling post-training is all they did. The result is a roughly 750B parameter model that lands at the frontier of agentic coding, a third the size of Moonshot's Kimi K3.

Steep jumps, not marginal gains

The published numbers are large deltas over GLM-5.2. From Together AI's model card:

• Terminal Bench 2.1 / 3.0: 88.2 / 28.3, with the 3.0 result up from 4.6 for GLM-5.2
• DeepSWE v1.1 (repository-scale engineering): 66.9, up from 46.2
• NL2Repo (repository generation): 58.0; FrontierSWE (open-ended multi-hour projects): 78.1
• SWE-Marathon v1.1 (ultra-long horizon): 42.5
• Humanity's Last Exam with tools: 62.5; Agents' Last Exam: 28.5
• CyberGym: 84.5 (up from 77.2); ExploitBench: 54.4 (up from 24.4)

Z.ai also claims a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench, reaching higher accuracy with fewer output tokens at every effort level. Treat all of this as vendor-reported until independent runs land, though the pattern (long-horizon, agentic, multi-tool) matches where post-training scaling is expected to pay off.

Where the engineering actually lives

According to the model card, Z.ai went well past coding katas into full-fledged expert workloads:

• Synthesized long-horizon environments with multi-step dependencies and hidden state, built by research agents from real task patterns and verified solvable by judge agents
• Verifiers synthesized without access to reference solutions, passing oracle, no-op, and unsolved-state checks before their binary rewards are used for training
• Carries over GLM-5.2's RL strategies, including SAO with compaction for long-horizon stability, with Z.ai reporting a 2.3x improvement in end-to-end RL training throughput from system optimizations

For full GLM-5.3, the recipe boils down to keeping the base, generating harder environments, hardening the verifiers so the model cannot game rewards, and pouring more RL compute at it. Training environments moved beyond coding exercises toward real units of expert work, some representing several days of an experienced engineer's effort, pushing the model to take ownership of substantial tasks end to end rather than relying on users to decompose and supervise each step.

Using it today

GLM-5.3 is live on Together AI's serverless endpoint under the model ID zai-org/GLM-5.3. Pricing on Together sits at $1.40 input, $0.26 cached input, and $4.40 output per 1M tokens, matching GLM-5.2's rates. Context window is 1M tokens, carried over from the 5.2 base.

A minimal call looks like:

from together import Together
client = Together()
response = client.chat.completions.create(
    model="zai-org/GLM-5.3",
    messages=[{"role": "user", "content": "Refactor this repo..."}]
)

Thinking is always on. You set reasoning_effort to low, high, or max per request, with max as the default and recommended setting for coding tasks. It is compatible with Claude Code, OpenCode, and other major coding agent harnesses out of the box.

One caveat: this is a gated release. Z.ai said plainly that GLM-5.3 is available now through GLM Coding Plan and ZCode, and that API access and open weights will be released in stages following rigorous safety evaluations. That is a departure from GLM-5.2's near-simultaneous weight drop.

Flash is a different animal

The tweet groups GLM-5.3 and GLM-5.3 Flash together, though they are architecturally different. GLM-5.3-Flash is neither an alias nor a compressed build of GLM-5.3. Full GLM-5.3 is a text-only post-training upgrade based on GLM-5.2. GLM-5.3-Flash starts from a newly trained base, adds native multimodal input, and has its own MIT-licensed official checkpoint.

Its architecture and training recipe were redesigned around capability and efficiency. For the first time in the GLM series, Z.ai introduced a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Concretely, of the 45 layers, 34 use linear attention, whose memory footprint stays flat regardless of conversation length, while the remaining 11 use sparse attention with a conventional KV cache that grows with every token.

At 320B total parameters and just 18B active, Flash outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. It is already on Hugging Face under MIT, and it is the multimodal path if your agent needs to look at screenshots or PDFs.

Emergent cyber capability

The other headline nobody buries in the release: this model finds real bugs. Post-training also developed strong vulnerability-discovery capability, which Z.ai has directed into an ongoing coordinated disclosure program with a public ledger, with 2,436 real vulnerabilities identified across 269 open-source projects.

That is the direct reason the weight release is gated. Lambert framed it as Z.ai's most capable model to date for cybersecurity tasks, delivering substantial improvements in vulnerability discovery, exploit analysis, and complex multistep security tasks.

What to update in your worldview

Two assumptions are due for revision. First, the ceiling on post-training-only gains is higher than most people were pricing in. If you can move DeepSWE from 46 to 67 without touching the base, the strategic value of environment libraries, verifier design, and RL infrastructure starts to look comparable to owning the pretraining run.

Second, the release cadence gap keeps widening. GLM ships roughly quarterly, with 8+ major releases in 18 months, the most consistent cadence among Chinese frontier labs. For teams picking a model to build agents on, that translates directly into how often your stack gets a free capability upgrade under a stable API surface.

For anyone building long-horizon coding agents, security tooling, or repo-scale automation, GLM-5.3 is worth wiring up as a comparison endpoint this week. If you need vision or the cheapest possible per-token cost, reach for Flash instead. The gap between them is smaller than the names imply.