# 📆 ThursdAI - Multimodal SOTAs (StepFun, Cohere), 3 new Qwens, Runway Alpha, Flux[Krea] &Charm agents

Source: https://www.youtube.com/watch?v=0kpTGI4WF2k
Recap page: https://rapidrecap.app/video/0kpTGI4WF2k
Generated: 2025-08-28T10:33:18.239+00:00

---
## Quick Overview

The AI news show "ThursdAI" covered a flurry of recent open-source model releases, with a significant focus on new iterations from Quen, including the Quen 3 Coder Flash, and updates to their reasoning and non-thinking models. Other key releases discussed were GLM 4.5 from ZAI (formerly Jipu AI), Command-R Vision from Cohere, and Step 3 from Stepfun, highlighting advancements in multimodal capabilities and efficiency, alongside a viral tool called Crush from Charm Bracelet.

**Key Points:**
- Quen released multiple new models, including the "Quen 3 Coder Flash" (also referred to as "Quen 3 30B A3B coder"), a three billion parameter model capable of running locally at high speeds, and updated their "thinking" and "non-thinking" models with 257 versions.
- ZAI (formerly Jipu AI) launched GLM 4.5, a 355 billion parameter model with unified reasoning and coding capabilities, though its benchmark reporting and comparison methods were critiqued for lack of transparency.
- Cohere released Command-R Vision, a state-of-the-art multimodal AI model for enterprises, featuring image and text reasoning, but with restrictive licensing (CC BY-NC) and compared primarily against Western models.
- Stepfun introduced Step 3, a new open-source multimodal reasoning model with 321 billion parameters and 38 billion active parameters, claiming significant speed improvements and new benchmark results, though access was initially limited to Chinese phone numbers.
- A viral tool named Crush from Charm Bracelet (a fork of Open Code) was highlighted for its extreme speed, processing "39 to 50 million tokens every hour," with the creator attributing the buzz partly to user dissatisfaction with Anthropic "nerfing" Claude.
- The show also touched on other releases like GLM 4.5's predecessor GLM 32B being strong in math and astronomy, RC's AFM and AFM 4.5 base models trained from scratch, and Alibaba's Tongyi releasing the open-source video generation model W1 2.2.
- The lack of expected major releases from OpenAI (like GPT-5) or an open-source equivalent was noted, with speculation about a secret model called "Horizon Alpha" on OpenRouter potentially being the latter.

**Context:** The video is an episode of "ThursdAI," a news and discussion show about recent developments in Artificial Intelligence, hosted by Alex Vulov with co-host Wolf from Raven Wolf. The episode, dated July 31st, focuses heavily on the rapid pace of open-source AI model releases that occurred in the week leading up to the broadcast, particularly within the last hour before the show. The discussion involves multiple guests, including Yam and N, who provide insights and analysis on these new models and tools.

## Detailed Analysis

This episode of "ThursdAI" provides a comprehensive overview of a particularly busy week in AI, marked by numerous open-source model releases. The hosts and guests express surprise at the sheer volume and speed of these developments, especially the continuous stream of updates from Quen, which included the Quen 3 Coder Flash, and enhanced reasoning and non-thinking models. GLM 4.5 from ZAI was presented as a strong contender, though its creators' benchmark reporting methods drew criticism for lacking transparency and selectively excluding top-performing models like Quen Coder. Cohere's Command-R Vision was highlighted as a significant multimodal offering for enterprises, emphasizing its enterprise-grade security and data integrity, but its non-commercial license was noted. Stepfun's Step 3 also emerged as a new multimodal reasoning model, with impressive technical specifications and benchmark claims, though initial access was restricted. Beyond model releases, the discussion covered the viral success of the Crush tool from Charm Bracelet, a high-throughput code processing utility, and touched upon other releases like RC's AFM models, Alibaba's video generation model, and the general trend towards multimodality in AI. The absence of expected major announcements from OpenAI, such as GPT-5, was also a point of discussion.

### Key Open Source Model Releases

- Quen 3 Coder Flash (3B params, fast local inference)
- GLM 4.5 (355B params, reasoning/coding)
- Command-R Vision (multimodal, enterprise-focused)
- Step 3 (multimodal reasoning, high speed)

### Quen Model Updates

- Introduction of Quen 3 Coder Flash, "breaking news" release
- Updates to "thinking" and "non-thinking" models (257 versions)
- Host notes six Quen releases in two weeks, suggesting staggered releases for exposure

### ZAI (Jipu AI) GLM 4.5 Analysis

- 355B parameter model with unified reasoning and coding
- Critiques on benchmark reporting: blended benchmarks, exclusion of Quen Coder from graphs, selective comparisons
- User "N" notes previous team's work on 1M context models and GLM 32B's strength in math/astronomy

### Cohere Command-R Vision Details

- State-of-the-art multimodal AI for enterprises
- Features image and text reasoning, enterprise-grade security
- Licensing restrictions (CC BY-NC, susceptible use policy)
- Benchmark comparisons criticized for cherry-picking and chart crimes

### Stepfun Step 3 Overview

- New open-source multimodal reasoning model (321B params, 38B active)
- Claims 4,000 tokens/sec/GPU speed, 70% faster than Deepseek v3
- Trained on 20T tokens (4T multimodal)
- Initial access limited to Chinese phone numbers; benchmark data presented in confusing charts

### Tools and Utilities

- Crush from Charm Bracelet (fork of Open Code) praised for extreme token processing (up to 50M/hr)
- Viral success attributed partly to "nerfing" of Anthropic's Claude
- Mention of "Horizon Alpha" as a potential open-source OpenAI model being tested

### Other Notable Mentions

- RC's AFM and AFM 4.5 base models (trained from scratch)
- Alibaba's Tongyi W1 2.2 (open-source video generation)
- Speculation on "Zenith" and "Summit" models as potential GPT-5 variants
- Lack of expected GPT-5 or open-source OpenAI release discussed