320 billion parameters, 18 billion on shift

When people called GLM-5.3-Flash fast this week, they were really talking about 18 of its 320 billion parameters. Z.ai confirmed that the mystery model everyone had been hammering on — filed under “ox-alpha” — is GLM-5.3-Flash: 320B total, 18B active per token, MIT-licensed, the first natively-multimodal release in the GLM-5 series, and one of the cheapest frontier-grade models on any rate card at launch. Public benches put it around 63 tok/s (field median ~90), an absurd number for a 320B model — and that’s the trick: it’s a Mixture-of-Experts, so only a slice of the weights fires per token, and here that slice is 18B. Once it’s resident, it answers like a model eighteen times smaller.

The catch is the first three words of that last sentence: once it’s resident. The footprint is the full 320B, and that gate is the rest of the story. BF16, 642.7 GB, fits on no single machine. The ladder down: a 512 GB unified-memory Mac holds 8-bit MLX (~334 GB) or Q8_0 GGUF (~341 GB) at near-full quality; a 128 GB workstation gets a 3-bit build (UD-IQ3_XXS), usable but degraded; NVFP4 (~198 GB) is Blackwell GPUs only; CPU-only lands in single-digit tok/s, which is batch work, not chat. And the first token at 1M context took ~34 s. The genuinely new thing is that nobody had to wait for tooling: GGUF, FP8-with-MTP, DFlash speculative-decoding and NVFP4 builds shipped within days of the weights. The gap between “released” and “runnable on your own machine” has nearly vanished — it just leaves the machine holding the whole thing as the only question.

The one you can actually run

Unless you own a 512 GB Mac, the week’s runnable release is Muse-Glimmer-30B — Meta’s Apache-2.0 distillation of the Muse Spark frontier line: ~29.6B dense, text+image in and text out, roughly 15–20 GB at 4-bit, so one Mac or one GPU. The card claims it beats Gemma4-31B and Qwen3.6-27B for its size on agentic work; that’s Meta’s own table, so it’s a claim. The numbers worth anchoring on are the speed figures, because they show what speculative decoding buys on hardware you have: with the card’s DFlash drafter, an RTX 5090 goes 74.9 → 233.4 tok/s, an M5 Max 26.6 → 50.2, an M4 Max 23.7 → 37.8. Same weights, same answers — 3.1× on the NVIDIA box, closer to 1.9× and 1.6× on the two Macs. A small drafter guesses several tokens ahead, the real model verifies in one pass; it’s the same trick making big models usable everywhere else this week.

The two Qwens are for different lives. Qwen3.8-27B (Apache-2.0) is the daily driver: 27B dense, ~16 GB at 4-bit — a 96 GB Mac, a single mid-range GPU or a 121 GB box all run it well — and the new best-overall local 27B, SWE-bench Pro 61.7. Expect low-to-mid double-digit tok/s; that’s labeled an estimate, since there’s no public measurement. Qwen3.8-Flash-Next is the Qwen4 dress rehearsal and the week’s cleverest architecture: 125B of model plus a 51B n-gram embedding table plus a 4B MTP head — with only 6B active per token, and the n-gram embeddings deliberately offloaded to host memory with async prefetch, so part of the model lives off the GPU. 262k context native, extendable to 1M; SWE-bench Pro 62.5; and an MLX 8-bit + MTP build already exists for Apple silicon. It fits a 512 GB machine at 8-bit; a 96–128 GB box runs a heavier quant.

The rest of the shelf, fast

DeepSeek-V4-Flash-Vision-Exp is the first vision model in the Flash line — 305B, MIT-licensed, weights posted ~10 days after the API debut, API-era limits included — needing a large machine or heavy quant, served through llama.cpp, vLLM or SGLang. Tencent Hy4-preview is a ~770B ternary MoE — every weight just −1, 0 or +1 — shrunk to ~229 GB under Apache-2.0; it wants 512 GB+ and custom Metal kernels that don’t exist yet, so watch it, don’t run it. The creative side doesn’t measure in tok/s at all: MiniMax-H3’s new accelerator LoRAs cut generation wall-clock for music (H3 itself runs on a big unified-memory Mac via a Metal port), ACE-Step XL got an MLX build for Apple silicon, and LTX-2.5 is the newer video model — all wall-clock-per-track or per-clip territory. And in agent-land, Pi reworked its session layer — durable storage, lane-scoped views, token tracking — which is a toolkit for people writing agents, not a model.

One pattern covers the whole week: total parameters are a memory bill, active parameters are the speed promise, and every interesting release cheated the first number with the second. Of the big ones, Muse-Glimmer-30B is what you can actually run right now; GLM-5.3-Flash and DeepSeek’s vision Flash want machines most readers don’t own; Qwen3.8-27B is the one to keep on call. The gap between “released” and “runnable” keeps shrinking — it just never mentions which machine.

Speeds are public benches or the labs’ own cards, labeled as such; sizing follows the published quant footprints; the one estimate is marked in the text.