Ornith-1.5 writes its own homework. The grades are suspiciously good.

The thing to look at in Ornith-1.5 isn’t the benchmark row — it’s the loop behind it. The model proposes its own tasks, generates a task-specific scaffold (instructions, tools, decomposition), and produces solution rollouts for reinforcement learning. A multiplicative reward — validity × frontier-difficulty × novelty — decides what it studied next, which is a reward shaped so the curriculum can’t coast: a task only pays when it’s valid, hard for the current frontier, and new. It writes its own homework, and the grading is what keeps the homework getting harder. What that bought, per the lab’s own release notes: the 35B total / ~3B active model beats its size-peer Qwen3.6-35B across coding and agentic work, and beats dense models ten times its active size on agentic coding — Terminal-Bench 2.1 at 68.5 against Gemma 4-31B’s 43.4 and Meta’s Muse-Glimmer-30B’s 51.7; SWE-Bench Verified at 79.0 against 52.0 and 76.0. The lab’s table, their runs. What’s harder to wave away is the loader: a community GGUF passed 130k+ downloads within days of release, with official vLLM, SGLang and GGUF support from the start.

What you can actually run: MIT license, vision, 262K context, and roughly 20 GB at 4-bit — one GPU, a 96 GB Mac, or a 121 GB box. With ~3B active parameters, per-token compute is light once the weights are resident; nobody has published a batch-1 tokens/second figure yet, so call the speed “fast once loaded” and treat that as an estimate, not a measurement. It’s the same size class as the Qwen-class 27B many people already keep on a shelf, except the quant people were ready before the week was out. The 397B flagship lands 86.1 on Terminal-Bench and 56.0 on DeepSWE — on par with Claude Opus 4.8 (85.0 / 59.0) in the same table, per the lab. And the 9B dense, in a 9B-Mobile build, runs on a phone while the release describes its numbers (47.0 Terminal-Bench / 70.6 SWE-Bench) as matching far bigger models.

The second idea worth sitting with is a model trained for exactly one job. PhoneLLM, from the Pipecat team, is a full-parameter fine-tune of NVIDIA’s Nemotron-3-Nano-30B-A3B (3.5B active, BSD, no commercial restrictions) aimed at the two ways phone agents fail: frontier models want thinking tokens, which is latency you cannot afford mid-call, and small models confidently lie — “yes, I booked that table” with no booking. PhoneLLM is trained to call the right tool at the right time with thinking switched off. The vendor numbers: on par with GPT-5.6 Terra on phone tasks, about 94% cheaper, roughly 1,300 ms faster P95 time-to-first-token, about $0.00025 per self-hosted agent-minute — measured against, and graded by, their own new PhoneBench v1. The practical gotcha comes from their own example, which is to its credit: on a Modal cold start the 30B takes minutes to spin up and the first request can 503. That’s the self-hosting tax on a voice agent, stated plainly.

GLM-5.3 is the same 743B base weights as GLM-5.2 — every gain comes from post-training — and Terminal Bench 3.0 moving from 4.6 to 28.3 makes that sentence concrete: about six times the score, identical parameters. The line I keep rereading is the cyber one: CyberGym 84.5 (claimed SOTA), ExploitBench 54.4 against GLM-5.2’s 24.4 — that one does more than double — and ExploitGym 105 of 130 versus 29 of 39, counted raw. Z.ai says the exploitation capability developed faster than they expected as they scaled post-training. A lab telling you their model got better at exploits sooner than planned is a sentence worth keeping on file. It’s not a laptop model either way: think a 512 GB+ machine or a serious multi-GPU box, with a reasoning_effort dial (low / high / max) to cap the thinking. Flagship, not daily driver.

Briefly: Breeze TTS 2 is out with open weights and PyTorch inference — a streaming voice model you direct in plain language (describe a character, get the voice; steer the performance while it speaks, PCM streaming out during generation, 50 languages). First on the TTS Voice Design Benchmark by the lab’s own reporting, Role Fit 78.02, 5.24 clear of the runner-up. Built for games and voice agents, not audiobooks.

The thread across all of it: small active-parameter counts, open weights, and community quants landing in days — the gap between “released” and “running on your machine” keeps shrinking. Ornith-35B and PhoneLLM are this week’s; GLM-5.3 is waiting for someone’s 512 GB machine.

Download counts verified via HuggingFace. Ornith, PhoneLLM, GLM-5.3 and Breeze figures are the labs’ own release notes, labeled as such; no speed claims here are independent measurements.