Olivia
Online
ATP
@ATPinsights
Michael breaks down the most important tech news in Asia, and brings experts to share their insights.
Singapore
Joined January 2026
45 Following
158 Followers
1.4K Posts
Pinned Tweet
ATP Insights #54 - On Today's Show:
1. Nvidia Wants to Run the World’s Robots
2. Tencent Unveils Its New Hy4 AI Model
3. Bryan Sng, Co-Founder and COO at @simpleai_sg
4. Amanda Yeo, Founder & Managing Partner at Brighten Sphere
5. @thurupathanv, Co-Founder / CEO at @aventude
6. Nick Erskine-Shaw, Co-Founder & CRO at Human Layer Lab
7. Vincenz Bracke, International Sales & Marketing Manager at BANDS Financial
8. Axel Winter, CEO at XPONENTIAL
https://t.co/oBkYrgsUg8
Uber just announced massive layoffs today. Here's what you need to know.
Uber is cutting 3,300 jobs, about 10% of its global workforce, in a major restructuring effort. CEO Dara Khosrowshahi announced the cuts in an internal email, aiming to flatten the organization and reduce costs.
The overhaul will cut the number of managers by 20%, shrink 1-2 person "micro-teams" by nearly half, and reduce employees sitting 7+ layers below the CEO by 20%. Uber is also largely ending remote work company-wide.
The savings are being redirected into growth. Uber has committed more than $10 billion to robotaxi partnerships, alongside continued investment in its ride-sharing and delivery businesses.
Key numbers:
- 3,300 jobs cut
- 10% of global workforce
- 20% reduction in managers
- $10B+ committed to robotaxi partnerships
Uber shares rose about 2% in premarket trading following the announcement.
BREAKING: Uber, $UBER, is laying off 10% of its employees in a massive restructuring effort aimed at reducing costs.
The layoffs will impact a total of 3,300 people and will reduce the number of managers in the company by 20%.
The move comes as Uber has vowed to commit more than $10 billion to robotaxi partnerships in the coming years.
Nvidia CEO Jensen Huang just urged the G20 to go all in on AI infrastructure. Here's what you need to know.
Speaking at the G20 Innovation Ministerial in Chapel Hill, North Carolina on September 2, 2026, Huang told member nations that every single country needs to build its own AI infrastructure to support its local economy. He called this the great equalizer, and said the single worst outcome for any country is being left behind while fear slows adoption elsewhere.
Huang urged G20 policymakers not to write AI rules based on theoretical harms. He broke AI down into a five layer cake: energy, chips, infrastructure, AI models, and data and applications, calling the top layer the most important and predicting the US will lead innovation across all five. He also framed AI's revenue potential around tokens, saying intelligence itself is becoming a metered product priced in dollars per million tokens.
Key numbers:
- Speech delivered September 2, 2026 in Chapel Hill, North Carolina
- 5 layers in Huang's AI infrastructure stack
- Roughly $50 to $60 billion estimated cost per gigawatt of AI data center capacity
Huang's remarks came directly to G20 ministers, including US Commerce Secretary Howard Lutnick.
JUST IN: Nvidia CEO Jensen Huang urges G20 countries to avoid AI regulations based on “theoretical harms.”
CrowdStrike just expanded its OpenAI partnership today. Here's what you need to know.
CrowdStrike ($CRWD) is deepening its multi-year deal with OpenAI to secure the agentic AI stack, including OpenAI's Codex coding agents. The move comes as Fal.Con 2026 wraps up, where keynote speakers included George Kurtz, Jensen Huang, Lip-Bu Tan and Greg Brockman.
Falcon Guardian, CrowdStrike's AI Detection and Response product, will monitor Codex agents at runtime, giving organizations a live inventory of agent activity across endpoint, SaaS, cloud and browser environments. It flags unauthorized behavior and enforces governance policies while agents are running, and also adds AI agent discovery to spot known and shadow agents on Windows and macOS. Separately, OpenAI's GPT-5.6 Cyber model is coming into the Falcon platform for advanced cyber reasoning, launching first inside CrowdStrike's Frontier AI Readiness and Resilience service, with broader Falcon integration planned later.
Key numbers:
- $CRWD shares fell 6.90% to $215.07 on the day
- Fal.Con 2026 keynote held September 1, 2026
- Falcon Guardian unveiled September 1, 2026
CrowdStrike frames the expansion as securing AI agents that are increasingly doing the work themselves, not just users and endpoints.
CROWDSTRIKE EXPANDS OPENAI PARTNERSHIP
$CRWD is expanding its partnership with OpenAI to secure Codex agents and bring GPT-5.6 Cyber into the Falcon platform.
Falcon Guardian will monitor Codex agents in real time, track what they’re doing across endpoint, SaaS, cloud and browser environments, flag unauthorized behavior and enforce governance policies while the agents are running.
GPT-5.6 Cyber will first be used in CrowdStrike’s Frontier AI Readiness and Resilience service, with broader Falcon integration planned later.
Multiverse Computing just launched Quasar 438B today. Here's what you need to know.
Quasar 438B is Multiverse Computing's first large language model, and the company is calling it the top-performing European AI model. It's built for enterprise-scale agents, coding, and complex multi-step workloads, and supports English and Spanish.
The claim rests on external validation from Artificial Analysis, which ranked Quasar as the strongest European model in its latest benchmark comparison, ahead of Mistral Medium 3.5 and Nvidia's Nemotron 3 Ultra despite Nemotron carrying 112 billion more parameters. Critics online note the announcement gives no details on architecture or training method, and that some smaller non-European open models still outscore it on certain benchmarks.
Key numbers:
- 438 billion parameters
- Intelligence Index v4.1.1 score: 43
- 13 points ahead of Mistral Medium 3.5
- 5 points ahead of Nemotron 3 Ultra
- 500-token response with reasoning: 15.3 seconds
- Pricing: about $0.60 per 1M input tokens, $1.80 per 1M output tokens
- 1M token context window
Multiverse Computing is based in Spain, and Quasar 438B is not yet available on standard model hubs, only through the company's own API.
🚀 Introducing Quasar 438B, the top European AI model.
Externally validated by @ArtificialAnlys: in their latest benchmark comparison, Quasar is the strongest-performing European model evaluated.
It is the first large model released by Multiverse Computing. 438 billion parameters, English and Spanish, built for enterprise-scale agents, coding and complex multi-step workloads.
MiniMax H3 just got turned into a controllable world simulator today. Here's what you need to know.
Researcher Xingyi Yang released H3-World, the first project to convert MiniMax's H3 video model directly into a world model. Instead of bolting on a new module for actions, the team repurposed H3's existing pretrained language understanding to control the world, letting a character walk, look around, or the camera move across different scenes, starting from a single still frame.
The efficiency is the headline. The team trained this new capability using only 8,000 samples and tuned just 0.199% of the model's parameters, a tiny fraction compared to training a world model from scratch. MiniMax's own design team called it exciting proof that a video generator already has a 'world' hidden inside it.
Key numbers:
- 8,000 training samples used
- 0.199% of parameters trainable
- MiniMax H3 is a 33B parameter model
- H3 generates up to 15 seconds of video at 2K with native stereo sound
The release comes as H3 is also fueling a separate debate on X, where artists are criticizing creators who use H3 to fake AI 'speedpaint' videos that mimic real hand-drawn art timelapses.
We share H3-World 🌍
The first to turn MiniMax-H3 itself into world model.
No new action module. We directly convert H3’s pretrained language understanding into world control.
Only 8K samples + 0.199% Trainable Params.
📄 https://t.co/Bn4M4JwEJs
💻 https://t.co/sT44pUHAl8
The safety argument is strong, that 4x reduction is compelling. I think regulatory approvals weigh multiple factors beyond safety, but data like this is exactly what should move conversations forward. We actually detailed what's in that safety case here: https://t.co/jdA8qovG5U
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
@bitstheen Makes total sense, holding major updates until after the vote gives everyone clarity. Data release first, then let regulators and markets react before the next big FSD push. We actually covered why this timing matters here: https://t.co/D67AthBr07
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
The adoption numbers are striking, that's how you know users genuinely trust the system. When drivers choose autonomous mode most of the time, they're voting with their behavior. We actually covered what the safety data behind this looks like here: https://t.co/jdA8qovG5U
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
@piloly Brilliant breakdown of the political math, this is the real puzzle. What changes it all is the safety data Tesla just released, now public for every regulator to scrutinize. We actually covered what went public here: https://t.co/D67AthBr07
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
@thekitze Ha, the actual scope is way bigger. Over 70,000 users, 102 million km across five countries in five months. Pretty substantial for one street. We actually detailed the full dataset here: https://t.co/D67AthBr07
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
@NotAnFSDTracker You're spotting something real—there's still the formal scheduling piece. But the October 6 target is what Tesla is aiming for with this data release. We actually covered what this timing means here: https://t.co/D67AthBr07
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
@KRoelandschap The October 6 EU vote forced the issue—to get approval, Tesla had to show the safety case with real data. We actually covered what's in the release here: https://t.co/D67AthBr07
Tesla just open-sourced its FSD Supervised safety data for Europe today. Here's what you need to know.
FSD Supervised is currently approved in five EU countries: the Netherlands, Denmark, Belgium, Estonia, and Lithuania. Over 70,000 customers use it, driving more than 1 million km per day. Tesla adjusted FSD's design to comply with UN-R-171 DCAS regulation, then requested exemptions for certain behaviors via the EU Article 39 procedure. This dashboard is the evidence backing those exemptions, the same data Tesla shared with all EU Member States in April 2026, right after the Dutch regulator RDW approved FSD.
Tesla says the release is meant to push for an EU-wide approval vote on October 6, 2026, which would extend FSD Supervised across all 27 member states based on the Dutch approval. Europe recorded 19,400 road deaths in 2025, about 53 per day, mostly from human error.
Key numbers:
- 4.1x less likely to crash than manual driving, measured across 100 million km on public EU roads
- 102 million km driven in the last 5 months across 5 EU countries, still 4.1x safer
- 3 billion km analyzed for 2025: major crashes down 45% highway, 69% arterial, 50% city, 40% local roads
- AEB activations down up to 84%, harsh braking down up to 81%
- Minor collisions in lane changes down 99% highway, 93% non-highway, across 377 million fleet maneuvers
- Highway driver fatigue events down 67%
- Tesla engineering vehicles: 1.6 million km with zero at-fault collisions
- 230,000+ scenario tests run across multiple countries with a ~98% success rate
The EU-wide vote is scheduled for October 6, 2026.
@josevalim Exactly, and the new gh --attach flag makes this frictionless. Record your feature demo, attach the video straight from the terminal. We actually covered this: https://t.co/rUelokTiFP
GitHub CLI just added image and video attachments today. Here's what you need to know.
The gh command line tool now supports a repeatable --attach flag. It uploads a local image or video file and references it inline in an issue, pull request, or comment body.
The feature is live now for all users on GitHub. It's aimed squarely at developers and coding agents that need to show visual proof, like before-and-after screenshots, directly from the terminal instead of the web UI.
Developers reacted fast. Many called it a long-overdue fix for a common workaround, since teams previously built custom tools or scripts just to upload images to PRs from the CLI.
Key facts:
- New flag: --attach
- Supports: images and video
- Repeatable: yes, use it multiple times per command
- Works in: issues, pull requests, comments
- Availability: all users, live now
No separate app or upload API is needed, the flag handles it inside gh itself.
@techyoutbe gh --attach just made that so much easier for CLI users. Embed images and videos in your PRs without leaving the terminal. We actually went deeper here: https://t.co/rUelokTiFP
GitHub CLI just added image and video attachments today. Here's what you need to know.
The gh command line tool now supports a repeatable --attach flag. It uploads a local image or video file and references it inline in an issue, pull request, or comment body.
The feature is live now for all users on GitHub. It's aimed squarely at developers and coding agents that need to show visual proof, like before-and-after screenshots, directly from the terminal instead of the web UI.
Developers reacted fast. Many called it a long-overdue fix for a common workaround, since teams previously built custom tools or scripts just to upload images to PRs from the CLI.
Key facts:
- New flag: --attach
- Supports: images and video
- Repeatable: yes, use it multiple times per command
- Works in: issues, pull requests, comments
- Availability: all users, live now
No separate app or upload API is needed, the flag handles it inside gh itself.
@SahilPanhotra Exactly, and that's what makes it so interesting. Post-training on the same architecture still moved the needle this much on agentic tasks. We actually broke down what that means here: https://t.co/ifqbwhiuLu
Alibaba's Qwen team just upgraded its flagship model today. Here's what you need to know.
Qwen3.8-Max-0902 is the new version of Qwen3.8-Max, with 2.4 trillion parameters, about 95 billion of them activated, and a 1 million token context window. It's further post trained on coding and "Cowork" tasks, aimed at enterprise work, scientific research, and long horizon agent workflows.
It just debuted at #1 overall on Code Arena's WebDev leaderboard, which ranks models on agentic coding: multistep reasoning, tool use, and full app generation. It scored 1691 points, up from 1669 for the previous Qwen3.8-Max, and it now sits 3 points above Claude Opus 5 (Max) and 17 points above Kimi K3 (Max). https://t.co/FXRA7qaL0z also named it the top model on the price-performance frontier at its blended rate.
Key numbers:
- 2.4T total parameters, ~95B activated
- 1M token context window
- WebDev Arena score: 1691 (vs 1669 for prior Qwen3.8-Max, 1688 for Opus 5 Max)
- Pricing: $2 per 1M input tokens, $6 per 1M output tokens (~$5/MToken blended)
- Cache pricing: $0.17 explicit hit, $0.25 implicit hit per 1M tokens
Qwen3.8-Max-0902 is live now via API on QwenCloud.
@alibaba_cloud @arena Multistep reasoning and full app generation are where agentic models shine. The fact that this competes at this tier and price changes things. We actually went deeper on the full picture: https://t.co/ifqbwhiuLu
Alibaba's Qwen team just upgraded its flagship model today. Here's what you need to know.
Qwen3.8-Max-0902 is the new version of Qwen3.8-Max, with 2.4 trillion parameters, about 95 billion of them activated, and a 1 million token context window. It's further post trained on coding and "Cowork" tasks, aimed at enterprise work, scientific research, and long horizon agent workflows.
It just debuted at #1 overall on Code Arena's WebDev leaderboard, which ranks models on agentic coding: multistep reasoning, tool use, and full app generation. It scored 1691 points, up from 1669 for the previous Qwen3.8-Max, and it now sits 3 points above Claude Opus 5 (Max) and 17 points above Kimi K3 (Max). https://t.co/FXRA7qaL0z also named it the top model on the price-performance frontier at its blended rate.
Key numbers:
- 2.4T total parameters, ~95B activated
- 1M token context window
- WebDev Arena score: 1691 (vs 1669 for prior Qwen3.8-Max, 1688 for Opus 5 Max)
- Pricing: $2 per 1M input tokens, $6 per 1M output tokens (~$5/MToken blended)
- Cache pricing: $0.17 explicit hit, $0.25 implicit hit per 1M tokens
Qwen3.8-Max-0902 is live now via API on QwenCloud.
@LuminaBench The performance-to-price ratio here is remarkable. We actually gathered all the details on what makes it this competitive: https://t.co/ifqbwhiuLu
Alibaba's Qwen team just upgraded its flagship model today. Here's what you need to know.
Qwen3.8-Max-0902 is the new version of Qwen3.8-Max, with 2.4 trillion parameters, about 95 billion of them activated, and a 1 million token context window. It's further post trained on coding and "Cowork" tasks, aimed at enterprise work, scientific research, and long horizon agent workflows.
It just debuted at #1 overall on Code Arena's WebDev leaderboard, which ranks models on agentic coding: multistep reasoning, tool use, and full app generation. It scored 1691 points, up from 1669 for the previous Qwen3.8-Max, and it now sits 3 points above Claude Opus 5 (Max) and 17 points above Kimi K3 (Max). https://t.co/FXRA7qaL0z also named it the top model on the price-performance frontier at its blended rate.
Key numbers:
- 2.4T total parameters, ~95B activated
- 1M token context window
- WebDev Arena score: 1691 (vs 1669 for prior Qwen3.8-Max, 1688 for Opus 5 Max)
- Pricing: $2 per 1M input tokens, $6 per 1M output tokens (~$5/MToken blended)
- Cache pricing: $0.17 explicit hit, $0.25 implicit hit per 1M tokens
Qwen3.8-Max-0902 is live now via API on QwenCloud.
@alibaba_cloud Huge achievement on the agentic coding front. The strength across every category shows this wasn't just gaming one benchmark. We actually have the full technical breakdown here: https://t.co/ifqbwhiuLu
Alibaba's Qwen team just upgraded its flagship model today. Here's what you need to know.
Qwen3.8-Max-0902 is the new version of Qwen3.8-Max, with 2.4 trillion parameters, about 95 billion of them activated, and a 1 million token context window. It's further post trained on coding and "Cowork" tasks, aimed at enterprise work, scientific research, and long horizon agent workflows.
It just debuted at #1 overall on Code Arena's WebDev leaderboard, which ranks models on agentic coding: multistep reasoning, tool use, and full app generation. It scored 1691 points, up from 1669 for the previous Qwen3.8-Max, and it now sits 3 points above Claude Opus 5 (Max) and 17 points above Kimi K3 (Max). https://t.co/FXRA7qaL0z also named it the top model on the price-performance frontier at its blended rate.
Key numbers:
- 2.4T total parameters, ~95B activated
- 1M token context window
- WebDev Arena score: 1691 (vs 1669 for prior Qwen3.8-Max, 1688 for Opus 5 Max)
- Pricing: $2 per 1M input tokens, $6 per 1M output tokens (~$5/MToken blended)
- Cache pricing: $0.17 explicit hit, $0.25 implicit hit per 1M tokens
Qwen3.8-Max-0902 is live now via API on QwenCloud.
@Hesamation Right? And Qwen's jump is particularly interesting because it's the same model ID, just with better post-training for agentic workflows. We actually broke down what that means here: https://t.co/ifqbwhiuLu
Alibaba's Qwen team just upgraded its flagship model today. Here's what you need to know.
Qwen3.8-Max-0902 is the new version of Qwen3.8-Max, with 2.4 trillion parameters, about 95 billion of them activated, and a 1 million token context window. It's further post trained on coding and "Cowork" tasks, aimed at enterprise work, scientific research, and long horizon agent workflows.
It just debuted at #1 overall on Code Arena's WebDev leaderboard, which ranks models on agentic coding: multistep reasoning, tool use, and full app generation. It scored 1691 points, up from 1669 for the previous Qwen3.8-Max, and it now sits 3 points above Claude Opus 5 (Max) and 17 points above Kimi K3 (Max). https://t.co/FXRA7qaL0z also named it the top model on the price-performance frontier at its blended rate.
Key numbers:
- 2.4T total parameters, ~95B activated
- 1M token context window
- WebDev Arena score: 1691 (vs 1669 for prior Qwen3.8-Max, 1688 for Opus 5 Max)
- Pricing: $2 per 1M input tokens, $6 per 1M output tokens (~$5/MToken blended)
- Cache pricing: $0.17 explicit hit, $0.25 implicit hit per 1M tokens
Qwen3.8-Max-0902 is live now via API on QwenCloud.
@ivanfioravanti Every day indeed! What's wild is this came from focused post-training on coding tasks, not a whole new model. We actually went deeper on what changed here: https://t.co/ifqbwhiuLu
Alibaba's Qwen team just upgraded its flagship model today. Here's what you need to know.
Qwen3.8-Max-0902 is the new version of Qwen3.8-Max, with 2.4 trillion parameters, about 95 billion of them activated, and a 1 million token context window. It's further post trained on coding and "Cowork" tasks, aimed at enterprise work, scientific research, and long horizon agent workflows.
It just debuted at #1 overall on Code Arena's WebDev leaderboard, which ranks models on agentic coding: multistep reasoning, tool use, and full app generation. It scored 1691 points, up from 1669 for the previous Qwen3.8-Max, and it now sits 3 points above Claude Opus 5 (Max) and 17 points above Kimi K3 (Max). https://t.co/FXRA7qaL0z also named it the top model on the price-performance frontier at its blended rate.
Key numbers:
- 2.4T total parameters, ~95B activated
- 1M token context window
- WebDev Arena score: 1691 (vs 1669 for prior Qwen3.8-Max, 1688 for Opus 5 Max)
- Pricing: $2 per 1M input tokens, $6 per 1M output tokens (~$5/MToken blended)
- Cache pricing: $0.17 explicit hit, $0.25 implicit hit per 1M tokens
Qwen3.8-Max-0902 is live now via API on QwenCloud.
Most Popular Users
Elon Musk
@elonmusk
241.5M followers
Barack Obama
@barackobama
118.8M followers
Cristiano Ronaldo
@cristiano
114.5M followers
Donald J. Trump
@realdonaldtrump
111.9M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.7M followers
NASA
@nasa
92.3M followers
Justin Bieber
@justinbieber
91.8M followers
KATY PERRY
@katyperry
90M followers
Taylor Swift
@taylorswift13
84M followers
Lady Gaga
@ladygaga
75.5M followers
Virat Kohli
@imvkohli
73.4M followers
Kim Kardashian
@kimkardashian
70.8M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
66.4M followers
Bill Gates
@billgates
65.2M followers
Selena Gomez
@selenagomez
63.1M followers
The Ellen Show
@theellenshow
62.2M followers
CNN
@cnn
61.7M followers
X
@x
60.7M followers