Sage-A1.5: a 35B open agent that leads its class on BrowseComp
Shanghai AI Lab's Apache-2.0 agent runs on a single GPU and tops sub-40B models on web and code tasks — as Reflection's 501B Beam lands on Hugging Face.
Copy markdown
A 35B open agent that runs on one GPU
Sage-A1.5 is a 35B mixture-of-experts model (~3B active) from Shanghai AI Lab, Apache 2.0, with a 256K context extendable to 512K. It leads sub-40B models on BrowseComp (81.7, or 85.5 with its built-in verification protocol) and SWE-Bench Pro (61.2), and ships FP8 and 4-bit builds that run on a single A100/MI300X or an Apple M-series Mac via vLLM, SGLang, llama.cpp, Ollama or LM Studio. For you: a near-frontier agent for long-horizon web research and coding you can self-host instead of rent.
Reflection's 501B Beam is now a download
Reflection AI's Beam-1.0 — a 501B sparse MoE with 23.4B active params and a 1M-token context — is on Hugging Face in BF16 (~1 TB), FP8 and NVFP4 under Apache 2.0. Reflection reports SWE-bench Verified 81.6, Terminal-Bench 2.1 80.1 and AIME 2026 97.7, and bills it as the 'Western open-weight frontier.' For you: a frontier-class open coder you can quantize and serve locally with llama.cpp, Ollama or LM Studio, or hit through Reflection's OpenAI-compatible API.
Mistral's Voxtral 3 makes voice agents cheap
Mistral launched Voxtral 3, an open-weight speech pipeline: a Transcribe STT model, a TTS engine with 30-minute voice cloning, and a voice-agents API that streams real-time transcription, reasoning and spoken replies across 35+ languages. API access is expected around $0.001/minute — roughly six cents an hour. For you: wire up a talking agent that acts on tasks, on self-hostable weights, at a fraction of hosted-voice prices.