The model that teaches itself

The lead: Ornith-1.5-35B-A3B

What it is. Ornith-1.5 is the new flagship from Ornith (built on Qwen3.5 / Gemma 4 with continued pretraining), released in three scales: a 397B MoE, a 35B MoE, and a 9B dense (with a quantized 9B-Mobile build that runs on phones). The headline is the 35B: 35B total, ~3B active per token, 262K context, vision, MIT-licensed, with official vLLM, SGLang and GGUF support — and a community GGUF already sitting at 130k+ downloads.

Intelligence. The interesting part isn’t the size, it’s the training loop. Ornith-1.5 extends a self-scaffolding idea into a closed self-improvement cycle: the model proposes new tasks, generates a task-specific scaffold (instructions, tools, decomposition), and produces solution rollouts for reinforcement learning — then a multiplicative reward (validity × frontier-difficulty × novelty) pushes it to keep generating harder, more informative tasks. It writes its own curriculum. Results: the 35B beats its size-peer Qwen3.6-35B across coding and agentic work, and despite only ~3B active it beats dense models — Gemma 4-31B and Meta’s Muse-Glimmer-30B — by wide margins on agentic coding (Terminal-Bench 2.1: 68.5 vs 43.4 and 51.7; SWE-Bench Verified: 79.0 vs 52.0 and 76.0). The flagship 397B lands 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, on par with Claude Opus 4.8 (85.0 / 59.0) while beating GLM-5.2 and DeepSeek-V4-Flash-0731. The 9B hits 47.0 Terminal-Bench 2.1 / 70.6 SWE-Bench — matching far bigger models.

Speed / ease. With ~3B active, per-token compute is light once resident; GGUF builds are already out, so the footprint is the gate. The 35B at 4-bit is roughly ~20 GB — it runs on a single GPU, a 96 GB Mac, or a 121 GB box. No public batch-1 tok/s figure I can cite, so treat speed as “fast once loaded” (estimate, not a measurement). The 9B-Mobile is the one that lands on a phone.

The rest, graded

PhoneLLM Alpha 1 — the open voice-agent LLM

What it is. From the Pipecat team: a full-parameter fine-tune of NVIDIA’s Nemotron-3-Nano-30B-A3B (30B total, 3.5B active, MoE), trained with NeMo and released under BSD with no commercial restrictions. It’s purpose-built for phone voice agents — inbound customer service, healthcare, retail, hospitality — paired with transcription and TTS through a framework like Pipecat.

Intelligence. The point is tool calling without thinking. Frontier models want thinking tokens enabled, which means long delays; small models often say “yes, I booked that table” without actually doing it. PhoneLLM is trained to call the right tool at the right time with thinking off. Vendor-reported: on par with GPT-5.6 Terra on phone tasks, ~94% cheaper, ~1,300 ms faster P95 time-to-first-token, ~$0.00025 per self-hosted agent-minute — plus a new PhoneBench v1 benchmark that scores accuracy, speaking style, latency, and per-minute cost.

Speed / ease. 30B A3B is very runnable: 3.5B active keeps per-token work light, and it’s BSD so you host it yourself. The Pipecat example runs it as a Modal endpoint — note the cold-start: a 30B spins up in minutes and the first request can return 503 until it’s resident. That’s the trade for self-hosting a voice agent.

GLM-5.3 — the flagship that doubled on cyber

What it is. Z.ai’s GLM-5.3 uses the same 743B base as GLM-5.2 — every gain comes from post-training. It’s a different release from the GLM-5.3-Flash covered last week: this is the dense flagship, and it’s the most capable open-weights coding model right now, with open weights and local serving support (SGLang, vLLM, TokenSpeed, Transformers, KTransformers CPU+GPU offload, Unsloth, Ascend NPU).

Intelligence. ~50% improvement over GLM-5.2 on Z.ai’s Code Bench, plus open-source SOTA on Terminal Bench 3.0 (28.3 vs GLM-5.2’s 4.6) and Agents’ Last Exam (28.5). The eyebrow-raiser is emergent cyber capability: CyberGym SOTA (84.5) and it more than doubles GLM-5.2 on exploitation benchmarks (ExploitGym 105/130 vs 29/39; ExploitBench 54.4 vs 24.4). Z.ai says it developed faster than they expected as they scaled post-training — worth noting with care.

Speed / ease. 743B is not a laptop model. It’s a MoE (GLM MoE DSA), and you can control the thinking budget via reasoning_effort (low/high/max, default max). Realistic sizing: a 512 GB+ machine or a serious multi-GPU box; KTransformers lets you split across CPU+GPU but it’s still a big footprint. This is a flagship, not a daily driver — but it’s the top of the open-coding ladder.

Breeze TTS 2 — a voice you can direct

What it is. BreezeBlue open-sourced Breeze TTS 2 weights and the PyTorch inference code (Aug 25). It’s a speech model built for real-time voice AI: natural-language voice design (describe a character, get a voice), voice direction (adapt a performance as intent changes), low-latency streaming (PCM streams while the model generates), and multilingual speech across 50 languages.

Intelligence. It ranks #1 on the TTS Voice Design Benchmark — Role Fit 78.02, Voice Diversity 708, 98.5% transcript pass rate — leading the next-best model (MiMo-V2.5-TTS) by 5.24 points and producing ~39% more distinct voices. The lab also released open benchmarks for Voice Design, Voice Direction, and Latency.

Speed / ease. The point is latency, not tok/s: it streams audio while generating, so a character can talk back in near-real-time. Weights + PyTorch inference are open, so it runs locally; the 50-language + accent-control angle makes it a fit for games, interactive stories, and voice agents rather than one-shot narration.

The takeaway

Four genuinely different flavors of new this week: a model that trains itself (Ornith-1.5), a model trained for one job — answering the phone without thinking (PhoneLLM), a flagship whose post-training unlocked capabilities nobody planned on (GLM-5.3), and a voice model built for interaction rather than narration (Breeze TTS 2). The common thread: small active-parameter counts plus open weights plus fast community builds — so the gap between “released” and “runnable on your own machine” keeps shrinking. Ornith-1.5-35B and PhoneLLM are the ones you can actually run this week; GLM-5.3 is the one that needs a serious box.

Credits

• Ornith — Ornith-1.5 (MIT); 35B MoE benchmarks from the release notes; community GGUF (Tiel-Coder) on HuggingFace.
• Pipecat / Daily — PhoneLLM Alpha 1 (BSD); vendor-reported figures from the announcement.
• Z.ai — GLM-5.3 (open weights); benchmarks from the model card.
• BreezeBlue — Breeze TTS 2 (open weights + PyTorch inference); benchmarks from the release.

Sources: HuggingFace model cards (verified via the HF API) and the labs’ release notes. Speeds are measured where a benchmark is cited; where not, they’re labeled estimates. No fabricated numbers.