DeAI Daily Brief: catching up on October's open-model releases

Today in DeAI: Meta's downloadable multimodal model, Reflection's shipped Beam weights, and new local-agent and speech options. This is a catch-up edition for material collected October 14–20; the underlying announcements did not all happen today. The common deployment question is what the released artifacts let a builder actually run.

Key facts

• Muse Spark Open has approximately 816B total parameters and 42B active per token; its expert set still requires substantial memory.
• Beam's integration guide now links Apache 2.0 checkpoints in BF16, FP8, and NVFP4, following its earlier announcement-only phase.
• Drama 3's community license has a $500,000 annual-revenue threshold for covered commercial use and separate hosted-service rules.

Meta's open multimodal model brings a large memory bill

Muse Spark Open is downloadable, but its active count is not a sizing guide. Hugging Face documents text, image, audio, and video inputs, with different modality support across serving backends. Its integration guide describes a BF16 checkpoint of about 1.5 TB and local quantized configurations that still need high-memory hardware. Why it matters: weight access creates a self-hosting option, but loading the model and serving its maximum context are separate capacity questions. Read DeAI's memory and serving report.

Beam moves beyond the waitlist

Reflection's weights now exist as deployable artifacts. The Hugging Face release guide links checkpoints, runtime integrations, and training recipes for the 501B model, resolving the availability question in DeAI's earlier launch coverage. Its performance table is attributed to Reflection and does not independently establish the launch's efficiency claims. Why it matters: developers can now pin a checkpoint and test those claims on their own workloads. Read the weights-release follow-up.

Liquid AI targets reasoning agents on owned hardware

LFM2.5-24B-A2B adds a smaller local-agent candidate. Liquid AI's model card lists 24B total and 2.45B active parameters, with a 131,072-token context and model variants for different runtimes. Its reported device speed and favorable tool-use scores remain vendor measurements, while its own comparison shows larger models ahead on some coding tasks. Why it matters: local inference is becoming easier to evaluate inside existing harnesses, but tool traffic and logging still determine the wider privacy boundary. Read the local-agent report.

Drama 3's editing workflow comes with specific permissions

Fish Audio adds speech revision and infilling, not unrestricted hosting rights. The Drama 3 card documents downloadable weights and local serving, while the license separately defines revenue eligibility, hosted-service restrictions, and exceptions. A small company's hosted speech API does not qualify merely because its revenue is low. Why it matters: a technically workable deployment may still need a separate commercial agreement for its actual product shape. Read the license and serving analysis.

DeepSeek's update needs a compatible encoding path

Flash-1017 reports agent gains and supplies explicit prompt tooling. DeepSeek's card reports DeepSWE rising to 77.0 from 74.2 for the predecessor and links MIT weights, but says the release has no Jinja chat template. Its reference encoder and parsing toolkit handle conversation and tool formats, leaving inference and execution to the caller. Why it matters: switching a model name without checking the backend's encoder can prevent an application from realizing the reported improvements. Read the checkpoint-update report.

Watching tomorrow: whether real deployments reproduce these published results, alongside the newly documented FLUX.3 Image Klein serving path and Apertus 1.5 training and evaluation artifacts.