AI Just Got Bigger & More Scalable!
This week brought a wave of openness across the AI industry, with Alibaba committing to open model weights at an unprecedented scale, DeepSeek delivering a cheaper model that outperforms its pricier sibling, Black Forest Labs unifying video and audio generation into a single step, and India's Sarvam AI inviting the public into its government backed AI mission. Here is what happened this week.
This Week at a Glance
- Foundation Models: Alibaba releases Qwen3.8-Max, its largest model yet. The model accepts text, images, and video as inputs, costs $2 per million input tokens, and open weights are expected to be released next week.
- Developer AI: DeepSeek moves V4-Flash into public beta after retraining it to outperform V4-Pro across all nine agent and coding benchmarks published by DeepSeek.
- AI Video: Black Forest Labs launches FLUX 3, its first video generation model capable of creating up to 20-second video clips with synchronized audio from a single unified architecture.
- Open Source: Indian AI company Sarvam launches Sarvam Circle, a public collaboration initiative under the government- backed IndiaAI Mission, with plans for weekly open-source model releases.
01 · Foundation Models
Alibaba Launches Qwen3.8-Max, Its Largest Open Model Yet, Priced at $2 Per Million Input Tokens
Alibaba · Qwen3.8-Max · August 2026
Released on August 3, 2026, Qwen3.8-Max is Alibaba's largest AI model to date, built to read
and process text, images, and video in a single session of up to one million words or the
rough equivalent in other media. The model is available immediately through Alibaba's cloud
service at $2 per million input tokens, placing it among the more affordably priced
multimodal offerings at this scale. On the text Arena leaderboard it ranks as the highest
scoring Chinese model, though it still trails several Anthropic models. Alibaba has
committed to releasing a freely downloadable version the following week, which would mark
the first time the company has made a model of this size available for anyone to download
and run independently.
Our Takeaway: Enterprise software teams that need a single model to process
documents, images, and video together now have a competitively priced option they can
self host or run through the cloud. That choice removes the dependency on any single
vendor and gives teams direct control over where their data goes.
02 · Developer AI
DeepSeek Releases a Retrained V4-Flash That Outperforms Its Own More Expensive Pro Model
DeepSeek · V4-Flash · August 2026
DeepSeek moved V4-Flash into public beta on July 31, 2026, releasing a version with entirely
redone post training that now scores higher than the larger and more expensive V4-Pro-
Preview across all nine agent and coding benchmarks the company published. The
underlying design and pricing of V4-Flash are unchanged from its earlier release. DeepSeek
has kept V4-Pro on a preview label, with a full official launch promised in the near future.
Our Takeaway: Development teams running automated coding or multi step task agents
at scale can now access scores close to top performing models at a fraction of the cost, with
no changes needed to existing code or workflows. The practical shift is that production
quality agentic workloads no longer require premium model pricing to be competitive on
output quality.
03 · AI Video
Black Forest Labs Launches FLUX 3, Generating Video and Synchronized Audio Together From a Single Model
Black Forest Labs · FLUX 3 · August 2026
FLUX 3 launched on July 23, 2026, as the first model from Black Forest Labs trained on
images, video, and audio within one shared architecture rather than combining separate
specialist models after the fact. It generates video clips up to 20 seconds long with audio that
is created in the same pass, meaning sound is produced alongside the picture rather than
added afterward. The model accepts text, still images, or existing video as its starting point.
Black Forest Labs is also applying the same architecture to robot movement prediction
through a partnership called FLUX-mimic, which has been tested on Audi production lines.
Video is currently available through a gated early access request at bfl.ai, with image
generation and a freely downloadable version planned for later in 2026.
Our Takeaway: Video producers and content studios that currently handle audio as a
separate post production step can now generate picture and sound together in one pass,
removing the manual work of syncing footsteps, dialogue, and ambient sound to finished
footage. Early access is open by request before the tool reaches general availability, giving
studios a window to test whether it fits into their production pipeline ahead of a public
launch.
04 · Open Source
Sarvam AI Opens Sarvam Circle to Developers and Institutions Building AI for India
Sarvam AI · Sarvam Circle · August 2026
Sarvam AI announced Sarvam Circle, a structured program that invites individual
developers, researchers, and institutions to contribute to the building of foundational AI
models for India. The program runs alongside Sarvam's role in the Indian government's
IndiaAI Mission, under which it is training large language models from scratch using
government provided computing resources. All models produced through the mission will be
released as open source under permissive licenses. The team publishes weekly model
updates and technical reports to keep contributors and the broader public informed of
progress.
Our Takeaway: Indian developers and enterprises that need AI tools designed for local
languages and regulatory conditions now have a structured, publicly funded path to
contribute to and build on top of those tools rather than adapting foreign systems not
designed for Indian contexts. The open source license means any business or researcher
can take the resulting models and build on them without negotiating proprietary access.
Across four different announcements this week, the clearest pattern is a shift toward
openness: downloadable model weights, public betas, government funded open source
pipelines, and gated early access replacing closed launches. Companies are competing less on
secrecy and more on performance per dollar and ease of adoption. For businesses evaluating
AI vendors right now, the practical result is more leverage and fewer reasons to accept a
single vendor lock in.
ASR AI Studio