The full stack goes open: giant weights on serverless, a 30B agent on one GPU, open RISC-V, one x86

816B weights on serverless

Meta's Muse Spark Open is now available on Fireworks' US serverless platform. Fireworks describes an 816B-parameter mixture-of-experts model with 42B parameters active per token and a 1M-token context window, built for long-horizon coding and agentic workflows — and says it runs with zero data retention by default. Fireworks pitches it for regulated professional tasks on the strength of the legal and medical benchmark results below. m:…3442

The benchmark sheet Fireworks published is aggressive: 91.1% on AIME26, 89.7% on GPQA Diamond, and 84.3% on SWE-bench Verified, per the company's post. In Meta's evaluations it scores 82.9% on Terminal-Bench 2.1 — though both sets of figures are vendor-reported. m:…3442

On professional tasks, Fireworks says the model ranks first of 75 models on Vals AI's Harvey Egal legal-agent benchmark at 25.4%, scores 53.3% on Rogo's Big Finance Bench — the top open-weights result — and hits 90.1% on Vals MedScribe. Fireworks says the model outperformed Claude Opus 5.5 on Rogo's Big Finance Bench. Its hallucination rate on AA-Omniscience is 33%, the company says. m:…3442

For regulated industries, the combination Fireworks is selling is open weights plus zero data retention by default — the compliance story that goes with the benchmark sheet. m:…3442

A 30B agent that fits on one GPU

Muse Glimmer 30B, described in a paper this week, is a 30-billion-parameter open-source causal language model engineered for agentic tasks on consumer hardware. With a perception encoder and 4-bit quantization, it performs multi-step reasoning and tool use locally on a single GPU — no cloud required — and is released under the Apache 2.0 license. m:…7060

The authors claim state-of-the-art performance for its size class on agentic benchmarks, including tool use, multi-step planning, and task completion ahead of Gemma4-31B and Qwen3.6-27B. Its perception encoder supports failure recovery locally, keeping the full agent loop on-device. Preprint claims, but the direction is clear: the agent stack is compressing onto hardware people already own. m:…7060

Open silicon

Google Research says it has completed a 32-bit open-source Coral NPU RISC-V architecture with a Matrix Execution Unit, delivering up to 1 TOPS while targeting 10 mW average power for edge inference. Alongside it came a v1.0 open-source IREE/MLIR compiler that takes models from JAX, PyTorch, and TensorFlow straight to Coral NPU hardware without custom assembly. m:…5644

Developers can prototype on the Synaptics Astra SR200 MCU series, and Coralboard devkits are globally available, the company says. The pitch: portable, energy-viable intelligence from model to silicon, with no proprietary toolchain in the middle. m:…5644

One x86

AMD and Intel are combining their x86 CPU optimization work for PyTorch in oneDNN, the open-source deep learning library governed by the Unified Acceleration (UXL) Foundation — following a recommendation from the x86 Ecosystem Advisory Group. AMD is integrating its ZenDNN optimizations into oneDNN and has joined the UXL Foundation as a steering member, the companies said. m:…1468

They are also collaborating on AI Compute Extensions, a next-generation x86 instruction set architecture designed to accelerate matrix multiplication, with PyTorch integration planned through oneDNN updates. The stated aim is to reduce fragmentation in the x86 AI ecosystem so PyTorch applications are easier to build, maintain, and deploy consistently across x86 systems. m:…1468