The last week of July and first days of August brought more model launches than most entire months of 2025. Here’s what matters — and what doesn’t yet.
The headline: Qwen 3.8-Max
Alibaba launched Qwen3.8-Max on August 3 — a 2.4T-parameter sparse MoE (95B active) that tops several agentic benchmarks at $2/$6 per million tokens. The standout numbers: OSWorld-Verified 86.1 (ahead of Fable 5’s 85.0), PaperBench 93.0 (highest reported), TerminalBench 86.6.
The catch: SWE-bench Pro 67.7 trails Fable 5 by 12+ points. And the “open weights next week” promise hasn’t materialized yet. If it does, this reshapes the open-weight landscape overnight.
Video generation: three new entrants in one week
MiniMax H3 — the first open-weight video model that matters
MiniMax H3 (July 31, weights August 3) is a 33B dense transformer generating 4–15 second video with native stereo audio in one pass. Open weights under MiniMax’s community license. The caveat: the 2K upscaling module is API-only; local inference caps at 768p.
FLUX 3 Video goes GA
FLUX 3 Video became generally available on August 4 with published pricing — Draft HD at $0.06/sec, standard HD at $0.17/sec, Full HD at $0.29/sec. Up to 20 seconds with native audio. BFL claims #1 Elo on its own T2V and I2V tests (ahead of Seedance 2.0 and MiniMax H3).
xAI Imagine Video 1.5 adds references and 1080p
Grok Imagine Video 1.5 got a major July 31 update: text-to-video, native 1080p, and image/voice references for character consistency. At $0.08/sec, it’s competitively priced and one of few models offering native full-HD.
Also this week
| What | Who | When | Why it matters |
|---|---|---|---|
| Gemini Robotics 2 | Google DeepMind | Jul 30 | Whole-body humanoid control + multi-robot collaboration. Research stage — published success rates are refreshingly honest (45.7% floor pick). |
| SeedRealtime | ByteDance | Aug 5 | Audio-visual full-duplex LLM in the Doubao app — watch, listen and speak simultaneously. No public API yet. |
| Shieldstral | Mistral | Aug 4 | 3B open-weights safety classifier (Apache 2.0) that matches models 7× its size. Policy-adaptive — accepts plain-language content rules at inference time. |
| Scribe v2 Realtime | ElevenLabs | Jul 31 | 150ms latency speech-to-text across 90+ languages. |
| Alpamayo 2 Super | NVIDIA | Aug 4 | 34B open-weights VLA model for autonomous vehicles. Niche but state-of-the-art in AV. |
The pattern
Three trends from this week:
- Omni-modal convergence. MiniMax H3, FLUX 3 Video and SeedRealtime all move toward single models that handle text, image, video and audio together — not bolted-on pipelines.
- Open weights are the competitive weapon. Qwen 3.8-Max promises to open a 2.4T model. MiniMax open-sourced H3. Mistral released Shieldstral as Apache 2.0. The race to release is accelerating.
- Video generation is the new battleground. Four significant video launches in one week (H3, FLUX 3 GA, Imagine Video 1.5 update, SeedRealtime) — a density not seen since the LLM wars of late 2024.
For model details, see each model’s page linked above. For the broader LLM landscape, see the 2026 LLM guide.