Cloudflare and Amazon ship open decision models to rival Jev
Two majors open-sourced tiny, fast agent routers the same day — plus a 27B model in 6GB, DeepSeek's CUDA-free Ascend kit, and Ling's not-yet-open weights.

Copy markdown
Clef turns agent routing into a 39ms call
Cloudflare open-sourced Clef (Qwen-27B, 64k context) and Clef-flash (Qwen-9B) under Apache 2.0 — decision models that classify and route instead of chatting, returning typed labels with probability scores. Clef-flash runs at 38.8ms median, Clef at 209.3ms, both beating Jev, and both pack a vision encoder Jev lacks. They're live on Workers AI and fully Jev-API-compatible, so you just swap the endpoint; an RL fine-tuning service is in preview.
Amazon's Strands Decider 2B is a self-hostable Jev killer
AWS's Strands Labs shipped Strands Decider 2B (Apache 2.0), a Qwen3.5-2B fine-tune that scores answer options — 'should this tool run?', 'which branch?' — in ~106ms median on an RTX 3090 (~150ms on an M3 MacBook) and hits ~72% on JevBench. There's no hosted API, so you bring the compute, but the weights, training data and scripts are all public — free and reproducible versus Jev's per-token API.
A 27B model that fits in 6GB is topping Hugging Face
prism-ml's Ternary-Bonsai-2-27B (Apache 2.0) squeezes a 27B model to 1.71 bits/weight — a 7–8.5GB download that holds ~94.6% of FP16 quality and runs ~26 tok/s on an M5 Pro laptop, with the full 262K context fitting in ~12.8GB using a 4-bit KV cache. The fresh MLX and GGUF builds are trending this week — a real weekend unlock for local coding and agents on a Mac or one modest GPU.
DeepSeek open-sources six Ascend tools to loosen CUDA's grip
DeepSeek, with Huawei, reportedly released six open tools — including TileLang (a CUDA alternative), DeepGEMM and FlashMLA — to run its stack on Huawei Ascend chips at claimed near-parity with Nvidia. If Nvidia supply is tight or you're targeting non-CUDA hardware, there's now an open path to try.
Ling-3.1-flash is announced, not open — don't plan around it yet
Ant Group's inclusionAI teased a 560B-total / ~25B-active MoE reasoner with up to 1M context, but there are still no weights, no license and no model card — just a 256K free trial in Ling Studio, a paid 1M tier after ~Oct 13, and open weights 'soon.' The benchmarks are vendor-reported and unverified; its predecessor took ~2 weeks from announce to weights, so treat today's numbers as a promise.