正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Today we’re announcing our “ThinkingCap” efficient model series with a 2× thinking token reduction on average in Qwen 3.6 27B, with up to 10x faster generation on individual examples.

Jaroslav Beck

18,221 subscribers

51,640 次观看 • 3 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

26 条评论

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

The results across seven out-of-domain benchmarks: 🏆 Up to 10x faster in specific responses 📉 ~46% fewer reasoning tokens on average 🎯 −0.7pp accuracy difference (comparable accuracy) 🔄 Fewer reasoning loops & dead-ends ⚡ Lower latency, lower inference cost, better real-world throughput

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

I am using this improved model in my local cluster “Bob” for some time now. I’s fast and also shortening the answers without degrading the output.

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

Our goal was to get to same answer sooner with fewer tokens burned and without loosing accuracy.

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

🔗 Full technical write-up with methodology & numbers in our blog post: 🔗 HuggingFace LINK:

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

PS: Efficiency not just in open models is our core thesis since we started year ago. If you want to work on frontier of efficiency, join us!

Martin Havel 的头像

Martin Havel3 个月前

Replicated on Czech-language RAG (tender docs, 3090, same quants): −29–55 % thinking, 0 quality loss, and 3× lower variance in thinking length across seeds — more predictable latency in production. Fits 24GB. Well done 👏

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

🔥

Tom Krcha 的头像

Tom Krcha3 个月前

🏎️ love it!

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

🙌🔥

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

;)

Chetin 的头像

Chetin3 个月前

Interesting.

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

your 🤖 Henry would love this model!

Ondřej Súkup 的头像

Ondřej Súkup3 个月前

Any plan for official mlx ?

Ddddddd 的头像

Ddddddd3 个月前

Qwen 3.6 36b-A3 will work also?

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

Our training algorithm will work for most open models.

Jan Kozák 🇨🇿🇪🇺🇺🇦🇵🇱 的头像

Jan Kozák 🇨🇿🇪🇺🇺🇦🇵🇱3 个月前

i will wait for mtp, this is currently slower

Gedeon 的头像

Gedeon3 个月前

Good stuff

Latent Node 的头像

Latent Node3 个月前

Great idea, similar to autothink - ?

Adel Bucetta 的头像

Adel Bucetta3 个月前

what makes the 2x token reduction so remarkable is it's a direct result of rethinking compute and dataflow from scratch

Pavol Zajicek 的头像

Pavol Zajicek3 个月前

Interesting 👍 Does it support MTP?

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

Yep

Pavol Zajicek 的头像

Pavol Zajicek3 个月前

Can I make Beat Saber 2 with it? 🙃✌️

Jaroslav Beck 的头像

Jaroslav Beck3 个月前

Go ahead! ;)

Johnny Hu 的头像

Johnny Hu3 个月前

Pitty it's at the beginning of summer holidays, tried to promote it to a community neewsfeed but everyone is gone for weeks :(. But thanks for making europe great again 😁 keep rocking

smeekky 的头像

smeekky3 个月前

Great job, @JaroslavBeck ! Keep up the excellent work. You are truly inspiring, and the Czech Republic appreciates what you do more than it may seem at first glance. Thank you to your entire team.

Václav Knapp 的头像

Václav Knapp3 个月前

Really cool guys!! 🇨🇿🇨🇿🔥 I have 4 questions below👇

相关视频

Managed to get QWEN 3.6 27B running in Cursor with localhost (no Ngrok) at 120 - 140 TPS on 2x 3090s.

Managed to get QWEN 3.6 27B running in Cursor with localhost (no Ngrok) at 120 - 140 TPS on 2x 3090s.

Youssof Al Toukhi

21,221 次观看 • 5 个月前

I built a J-Space explorer with several curated sessions. It shows all three regions described in Anthropic's paper on J-Space: sensory, workspace and motor. Model: Qwen 3.6 27B 4-bit

I built a J-Space explorer with several curated sessions. It shows all three regions described in Anthropic's paper on J-Space: sensory, workspace and motor. Model: Qwen 3.6 27B 4-bit

WeZZard

38,567 次观看 • 3 个月前

I put Qwen 3.6 27B on a B200 for the fun of it. I told Fable to apply my past optimisation principles. Result? 1000 TPS Qwen 3.6 27B NVFP4 (550 TPS average long decode) I used the same principles that got me 250 TPS on 2x 3090s. Now I need to buy a DGX Station 🤪.

I put Qwen 3.6 27B on a B200 for the fun of it. I told Fable to apply my past optimisation principles. Result? 1000 TPS Qwen 3.6 27B NVFP4 (550 TPS average long decode) I used the same principles that got me 250 TPS on 2x 3090s. Now I need to buy a DGX Station 🤪.

Youssof Al Toukhi

46,527 次观看 • 2 个月前

Vibe coding at 40,000 feet in the air at 75TPS+ on my MacBook with no WiFi. Qwen 3.6 27B on MTPLX V2 (coming soon!)

Vibe coding at 40,000 feet in the air at 75TPS+ on my MacBook with no WiFi. Qwen 3.6 27B on MTPLX V2 (coming soon!)

Youssof Al Toukhi

61,734 次观看 • 3 个月前

You can now run Qwen 3.6 27B on <10GB of memory and retain 90% of the intelligence. Today, PrismML dropped Bonsai 27B - which takes their unique compression approach they have already successfully applied to Qwen3.6 8B and Flux2-Klein and applied it to a properly smart model. Here is it using an entire computer for deep research. Long horizon agentic calls already are strong with Qwen3.6 27B and they still are here - just on 10x less memory.

You can now run Qwen 3.6 27B on <10GB of memory and retain 90% of the intelligence. Today, PrismML dropped Bonsai 27B - which takes their unique compression approach they have already successfully applied to Qwen3.6 8B and Flux2-Klein and applied it to a properly smart model. Here is it using an entire computer for deep research. Long horizon agentic calls already are strong with Qwen3.6 27B and they still are here - just on 10x less memory.

Tim Carambat

63,449 次观看 • 3 个月前

I tested Gemma 4 12b on my Macbook Pro M5 Max with 128GB of unified memory. It runs VERY slow for a 12b parameter model. 44 tokens/second. The lava lamp that it produced was also terrible. I would stick to Qwen 3.6 27b or Qwen 3.6 35b. Google is still way behind, even with their local models.

I tested Gemma 4 12b on my Macbook Pro M5 Max with 128GB of unified memory. It runs VERY slow for a 12b parameter model. 44 tokens/second. The lava lamp that it produced was also terrible. I would stick to Qwen 3.6 27b or Qwen 3.6 35b. Google is still way behind, even with their local models.

BridgeMind

70,626 次观看 • 4 个月前

72+ TPS on Qwen 3.6 27B on a Macbook pro M5 max. MTPLX V2 out now! The fastest way to run models on MLX.

72+ TPS on Qwen 3.6 27B on a Macbook pro M5 max. MTPLX V2 out now! The fastest way to run models on MLX.

Youssof Al Toukhi

90,489 次观看 • 3 个月前

LFM2-Audio just dropped! It's a 1.5B model that understands and generates both text and audio Inference 10x faster + quality on par with models 10x larger Available today on Hugging Face and our playground 🥳

LFM2-Audio just dropped! It's a 1.5B model that understands and generates both text and audio Inference 10x faster + quality on par with models 10x larger Available today on Hugging Face and our playground 🥳

Maxime Labonne

85,053 次观看 • 1 年前

100 TPS on 16 concurrent processes running a modified version of Cursor using QWEN 3.6 27B to build a 3D Bowmasters game.

100 TPS on 16 concurrent processes running a modified version of Cursor using QWEN 3.6 27B to build a 3D Bowmasters game.

Youssof Altoukhi

11,243 次观看 • 5 个月前

Announcing Coffee: build and iterate on your UI 10x faster with AI ☕️👇

Announcing Coffee: build and iterate on your UI 10x faster with AI ☕️👇

Coframe

364,341 次观看 • 2 年前

Qwen-Image-2.1 in 4 steps is here ⚡ ViggleAI distilled Qwen-Image-2.1 into a 4-step turbo model, 6× faster, and holds up side by side with the full model ▶️ on Spaces

Qwen-Image-2.1 in 4 steps is here ⚡ ViggleAI distilled Qwen-Image-2.1 into a 4-step turbo model, 6× faster, and holds up side by side with the full model ▶️ on Spaces

Hugging Apps

96,750 次观看 • 1 个月前

Qwen3.8-27b produces beautiful results very presentable 😍 Way better than 3.6-27b by a long shot. My three.js and glsl tests are going on, but here is few examples, same prompt with same default settings all run on my 3090. qwen3.6-27b-Q5 vs qwen3.8-27b-Q5

Qwen3.8-27b produces beautiful results very presentable 😍 Way better than 3.6-27b by a long shot. My three.js and glsl tests are going on, but here is few examples, same prompt with same default settings all run on my 3090. qwen3.6-27b-Q5 vs qwen3.8-27b-Q5

AJ

93,007 次观看 • 2 个月前

Today we’re introducing Navigator n2. It’s a frontier computer-use model, with just 27B parameters.

Today we’re introducing Navigator n2. It’s a frontier computer-use model, with just 27B parameters.

Devi Parikh

782,282 次观看 • 2 个月前

1/ Offline computer-use models are here. Today, one running on our Mac made it all the way to hotel checkout. We ran Qwen3.8-27B from Qwen through Cua Driver. It also turned a flyer into a Calendar event and reconciled receipts. Every model token stayed on the Mac 🧵

1/ Offline computer-use models are here. Today, one running on our Mac made it all the way to hotel checkout. We ran Qwen3.8-27B from Qwen through Cua Driver. It also turned a flyer into a Calendar event and reconciled receipts. Every model token stayed on the Mac 🧵

Francesco

27,540 次观看 • 2 个月前

Orinth-1.0-35b vs Qwen-3.6-27b I've asked both to build a single html file of Tetris game. First here's the Orinth 35b I think it did a fine job. Every control works, the game looks fine visually. How did Qwen 3.6 27b did? See post #2

Orinth-1.0-35b vs Qwen-3.6-27b I've asked both to build a single html file of Tetris game. First here's the Orinth 35b I think it did a fine job. Every control works, the game looks fine visually. How did Qwen 3.6 27b did? See post #2

Mia

62,696 次观看 • 4 个月前

Breaking: Browser Use is faster than a human⚡ Running Qwen 3.8 27B on 2x B200s with DFlash2. > speed ✅ > cost ✅ > accuracy ✅ (most of the time) Qwen 27B is better at using the internet than you are. The only bottleneck is website loading. Comment if you want us to deploy this to our cloud 👀

Breaking: Browser Use is faster than a human⚡ Running Qwen 3.8 27B on 2x B200s with DFlash2. > speed ✅ > cost ✅ > accuracy ✅ (most of the time) Qwen 27B is better at using the internet than you are. The only bottleneck is website loading. Comment if you want us to deploy this to our cloud 👀

Browser Use

1,107,892 次观看 • 2 个月前

A nice aspect of our new Gemini 3.6 Flash model is that it is much more token efficient than our 3.5 Flash model. Here's a side-by-side demonstration of that. Nice work by everyone who worked on this model and release!

A nice aspect of our new Gemini 3.6 Flash model is that it is much more token efficient than our 3.5 Flash model. Here's a side-by-side demonstration of that. Nice work by everyone who worked on this model and release!

Jeff Dean

212,506 次观看 • 3 个月前

Today, we’re announcing a $300 million Series C in a mix of primary and secondary funding at a $3.5 billion valuation. We’re excited to bring on @Sequoia as our lead investor in the round, with participation from Spark Capital, Marathon, COATUE, CRV, & a16z.

Today, we’re announcing a $300 million Series C in a mix of primary and secondary funding at a $3.5 billion valuation. We’re excited to bring on @Sequoia as our lead investor in the round, with participation from Spark Capital, Marathon, COATUE, CRV, & a16z.

Mercury

130,617 次观看 • 1 年前

Qwopus 3.6-27b Coder I had a lot of requests to test it, so I did. I ran the same tests I’ve done on other models. It scored lower on tool-eval-bench than both Qwen 3.6 27b and Qwen 3.6 35b. It also produced broken single-file Tetris and Solar System HTML demos, even after 3 fresh attempts. In the Tetris html, the game stalled when you completed a line. It also randomly changed the pieces when moving them left or right. In the Solar System html, clicking on the sun or any planet broke the rendering — the objects just disappeared. Conclusion: The regular Qwen 3.6 27b and 35b models performed better. What do you think of Qwopus 3.6-27b Coder? Full results & RAW files: Full prompts in the posts below.

Qwopus 3.6-27b Coder I had a lot of requests to test it, so I did. I ran the same tests I’ve done on other models. It scored lower on tool-eval-bench than both Qwen 3.6 27b and Qwen 3.6 35b. It also produced broken single-file Tetris and Solar System HTML demos, even after 3 fresh attempts. In the Tetris html, the game stalled when you completed a line. It also randomly changed the pieces when moving them left or right. In the Solar System html, clicking on the sun or any planet broke the rendering — the objects just disappeared. Conclusion: The regular Qwen 3.6 27b and 35b models performed better. What do you think of Qwopus 3.6-27b Coder? Full results & RAW files: Full prompts in the posts below.

Mia

11,183 次观看 • 4 个月前

Qwen Image 2.1 is here! 🖼️ A 7B params native image generation and editing model, with up to 10 image references The model comes with it's own prompt enhancement LLMs, integrated with diffusers 🧨 and ComfyUI ▶️ on Spaces

Qwen Image 2.1 is here! 🖼️ A 7B params native image generation and editing model, with up to 10 image references The model comes with it's own prompt enhancement LLMs, integrated with diffusers 🧨 and ComfyUI ▶️ on Spaces

Hugging Apps

92,424 次观看 • 1 个月前