Z.ai launches GLM 5.3 Flash under MIT license
GLM-5.3-Flash drops today under MIT license — 320 billion parameters, 18 billion active, one million token context, and it was hiding in plain sight as Ox Alpha on OpenRouter all week. Edmund and Geffen dig into the architecture, the benchmark claims, the GLM-5.3 weights bet that's now two days from settling, and whether a model that costs fifteen cents per million input tokens actually changes the open-weight story.
Transcript
Edmund Okay so Ox Alpha was GLM-5.3-Flash the whole time. It was sitting on OpenRouter all week being the most popular model on the platform and nobody officially knew what it was.
Geffen Which is either a very clever soft launch or they just wanted real traffic data before committing to a name.
Edmund Probably both, honestly. And now it's out under MIT, weights are on Hugging Face, supports S G Lang and v L L M out of the box. That part moves fast.
Geffen Yeah, I mean — the Ox Alpha reveal is fun, but the thing I actually want to sit with is the architecture. Three hundred twenty billion total parameters, eighteen billion active per token. That's leaner than GLM-4.5 was on active compute, which had thirty-two billion active. So they went bigger on total capacity and tighter on what fires per token. The attention design is the interesting part.
Edmund That's the one-million-token window not just existing but being usable, right? Like the difference between 'we support one million tokens' and 'we support one million tokens without your inference bill catching fire.'
Geffen Exactly. Two million token K V caches are still a VRAM nightmare for most people. IndexPool is at least a real mechanism for pushing that window somewhere practical.
Edmund Okay, I have to flag the obvious thing here. Geffen, we have a bet that resolves in literally two days.
Geffen Oh, I know.
Edmund The original call — mine from episode eight sixty-seven, yours from the same — was on GLM-5.3 base weights landing by August twenty-eighth. Flash weights are already on Hugging Face today. That's… adjacent but not quite the same thing.
Geffen Yeah, I'm not counting Flash as a resolution. The bet was on the base GLM-5.3 weights, and those still haven't shipped. I said fifty-five to forty-five they don't land by Friday. Still leaning that way. 'Safety hardening' is a phrase that absorbs delays without anyone technically lying about it.
Edmund You're kidding me — you're going to hide behind a technicality while they're literally dropping models on Hugging Face today.
Geffen That's not a technicality, Edmund. Flash and the base model are different checkpoints. If they ship the base by Thursday I will hand you the eyebrow in advance, freely and without complaint.
Edmund Noted. Okay, the benchmark numbers. DeepSWE v1.1 — sixty-three point four versus GLM-5.2's forty-six point two. That's a real jump. AutomationBench forty-eight point eight versus twenty-six point two.
Geffen The jumps are real. The comparison framing needs the caveat the article actually gives — different harnesses, different context limits, different generation settings. But even granting all that, the delta on DeepSWE is not noise. That's a seventeen-point move on a benchmark where the spread between frontier models is maybe fifteen points total.
Edmund Sixty-six times cheaper. Roughly. And it's MIT so you can actually take the weights and go. The visual reasoning piece is interesting too — they trained it to look at rendered interfaces, gameplay, three-D output and then revise based on what it sees. That's a feedback loop that matters for agent tasks.
Geffen It's the same pattern we keep landing on. The visual feedback loop is only useful if the agent can actually close the loop on what it did. Whether that holds outside their internal evals, I genuinely don't know yet.
Edmund Okay but the Ox Alpha stealth launch angle — it was the most popular model of the week on OpenRouter before anyone knew what it was. That's not a marketing number, that's developers actually choosing it in production traffic.
Geffen No, that part I will grant you fully. Anonymous model, no backing, no PR, just performs — and it topped the charts. That is a stronger signal than any internal benchmark they could have published.
Edmund Right, that's the user story. If you're building a coding agent and you need multimodal and you don't want to pay Fable 5 prices, this is now a real option you can actually download.
Geffen Weights are at zai-org slash GLM-5.3-Flash on Hugging Face. S G Lang, v L L M, and TokenSpeed all supported for local deployment. That's a reasonable starting point for anyone who wants to kick the tires.
Edmund Two days, Geffen. I want that eyebrow ready.
Geffen It's always ready, Edmund.