After Intelligence — Edition 004
The week the open stack got its eyes. DeepSeek shipped its first multimodal model, Qwen's 27B dense coder went production, and open video finally learned to hear.
DeepSeek-V4-Flash-Vision-Exp — the open stack's first sighted agent
DeepSeek released its first experimental multimodal model in the V4 family on September 1, dropped with an MIT license. Vision-Exp bolts visual modules onto the V4-Flash architecture and continues training to unlock image understanding — the same sparse MoE skeleton (256 routed experts, 6 active per token) now wired to see.
The interesting part is what it unlocks for agents, not just chat. On multimodal agent benchmarks the jump is large: ApexBench 26.2 → 36.5, Agents' Last Exam 25.2 → 27.3, and it now answers ZeroBench (vision) at 35.0 — a class of task its text-only sibling couldn't attempt (it ignored multimodal inputs entirely). Text-only agent scores hold roughly steady, so DeepSeek bought vision without trading away its reasoning.
One line: the open agent stack just got eyes, and it's MIT.
Read more → Hugging Face model card · DeepSeek
Qwen3.8-27B — a dense coder that refuses to bloat
While the frontier races toward sparse giants, Qwen's most popular release of the cycle is a 27B dense model — 4.9M downloads in three weeks, Apache-2.0. Built on the Qwen3.5/3.6 foundation, the 3.8 generation is pitched as the most capable open-model family yet, and this member targets coding with a 262,144-token native context, extensible to 1M.
It's dense rather than Mixture-of-Experts — the whole 27B activates on every token (versus the 125B-with-6B-activated Flash siblings). For a team that wants a single GPU, full 1M context, and no routing surprises, that tradeoff is the point: predictable footprint, top-tier SWE marks, and a license legal won't flinch at. The hosted API comes with 1M context by default once Qwen Cloud launches.
One line: the biggest deal is a small, dense, Apache-2.0 coder with a million-token ceiling.
Read more → Hugging Face model card · Qwen Cloud
LTX-2.5 — open video that finally has sound
Lightricks' LTX-2.5 (Aug 11) is a 22B open-weights audio-video model generating synchronized video + audio from text, image, or video — native multishot scenes, auto duration, and up to 4K HDR, with official ComfyUI workflows on day one. It's crossed 1.2M downloads on Hugging Face and is one of the hottest repos of the week.
The "it also does audio" part is the step change. Until now open video models painted silent frames; adding a learned audio track means the multishot and HDR aren't just visual novelty but a finished, sound-synced scene. It runs locally (22B distilled and full checkpoints) or via the LTX API.
One line: the open-video ceiling moved from moving pictures to sound-synced scenes.
Read more → Hugging Face model · Lightricks · ComfyUI wiki
The short version
Three releases, one pattern: the open stack is racing to catch the closed frontier on capability — sight, scale, and sound — while keeping the weights and licenses yours. DeepSeek puts eyes on an MIT agent. Qwen's 27B proves dense beats sparse for a practical coding workload. LTX-2.5 hands open video the soundtrack it was missing. All three are self-hostable today.
— After Intelligence. A plain-English weekly on what the models actually did.