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。

Concept Country has developed its first open-source model, Qwen Concept 27B. With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.

Concept

4,271 subscribers

64,875 次观看 • 5 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

9 条评论

Concept 的头像

Concept5 天前

The method comes from "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights" by four MIT CSAIL researchers, published November 2, 2026: - Rikiya Takehi (joint first author) - Jayden Teoh (joint first author) - Alex Zhang - Omar Khattab, whose group it is; he's also known for DSPy and ColBERT. This method immediately inspired us to work to solve a long-standing problem with Qwen models: thinking often outruns the context window. Would it be possible for us to directly edit the weights so a model thought less, while retaining its intelligence? The answer: yes. Overall, this reduces the tokens used on GPQA Diamond by 30% while retaining accuracy.

Concept 的头像

Concept5 天前

Using a custom mechanistic interpretability harness, we achieved this work entirely on local hardware: @NVIDIAAI DGX Spark GB10s handled layer investigation, rank-1 updates & evaluation loops. Replicating the MIT CSAIL team’s work with a model nearly 30x its size, we show that direct weight editing has more potential beyond just fixing loops.

Concept 的头像

Concept5 天前

With a third less deliberation, Qwen Concept 27B actually outperforms its base model.

Concept 的头像

Concept5 天前

In the coming days, we will provide BF16, FP8 and NVFP4 quantization for this model, which will be available for free at Make sure to follow @conceptcountry for updates.

Wbackdown 的头像

Wbackdown5 天前

54% less thinking is wild. Does it actually feel smarter in practice or just faster?

Brian Wesley 的头像

Brian Wesley5 天前

Cutting thinking tokens by 54% with a weight edit is a different claim from the MATH score jumps. Tokens per joule is the number worth repeating👍

Kevin Bondzio 的头像

Kevin Bondzio5 天前

Nice. One thing worth adding: a random edit of the same size as a control. In my runs it did nothing (97% → 97%), which is what convinced me the effect was real. Also check whether it survives a bit of fine-tuning. Mine didn't (20 steps).

Snorky!「🐬」的头像

Snorky!「🐬」5 天前

I'M CONCEIVING

elonsuxballs69 的头像

elonsuxballs695 天前

Are you hiring?

相关视频

Qwen 3.8 27b Makes BEUATIFUL websites!! Getting these on GH for people to look at as well. @Qwen and Qwen Developers this is such an amazing model for websites and components!

Qwen 3.8 27b Makes BEUATIFUL websites!! Getting these on GH for people to look at as well. @Qwen and Qwen Developers this is such an amazing model for websites and components!

Loktar 🇺🇸

44,975 次观看 • 2 个月前

Qwen 3.8 27b Pagoda test compared to 3.8 max and 3.6. BF16 results! Looks great!

Qwen 3.8 27b Pagoda test compared to 3.8 max and 3.6. BF16 results! Looks great!

Loktar 🇺🇸

36,732 次观看 • 2 个月前

Qwen 4 Unknown Preview bee on a sunflower. Qwen used significantly less tokens overall compared to Deepseek Flash and looks really good! Hope we can run this at home!

Qwen 4 Unknown Preview bee on a sunflower. Qwen used significantly less tokens overall compared to Deepseek Flash and looks really good! Hope we can run this at home!

Loktar 🇺🇸

13,007 次观看 • 4 天前

🤝 Arcee.ai, an American AI lab, built its frontier open model family on NVIDIA Blackwell Ultra. Its flagship MoE model, Trinity-Large-Thinking, was RL post-trained using NVIDIA NeMo open libraries. Arcee optimized agentic model inference with NVIDIA Dynamo and vLLM, along with NVIDIA accelerated networking, to deliver low token cost, achieving: ✅ 3T+ tokens served on OpenRouter in its first two months ✅ $0.90 per 1M output tokens ✅ 2nd on PinchBench for open agentic models Explore Arcee's open models, powered by NVIDIA ➡️

🤝 Arcee.ai, an American AI lab, built its frontier open model family on NVIDIA Blackwell Ultra. Its flagship MoE model, Trinity-Large-Thinking, was RL post-trained using NVIDIA NeMo open libraries. Arcee optimized agentic model inference with NVIDIA Dynamo and vLLM, along with NVIDIA accelerated networking, to deliver low token cost, achieving: ✅ 3T+ tokens served on OpenRouter in its first two months ✅ $0.90 per 1M output tokens ✅ 2nd on PinchBench for open agentic models Explore Arcee's open models, powered by NVIDIA ➡️

NVIDIA AI Infrastructure

27,377 次观看 • 3 个月前

There’s a good chance your open source model is costing more than the frontier. Cheap tokens ≠ cheap tasks. Here, we introduce Intelligence Density, and Density Aware Training, our post-training technique to achieve less wasted compute, better learning, all with no knobs to tune. Enabled by default in every Trajectory model.

There’s a good chance your open source model is costing more than the frontier. Cheap tokens ≠ cheap tasks. Here, we introduce Intelligence Density, and Density Aware Training, our post-training technique to achieve less wasted compute, better learning, all with no knobs to tune. Enabled by default in every Trajectory model.

Trajectory

58,101 次观看 • 1 个月前

Qwen 4 vs Deepseek Flash one shot Christmas Village. Qwen 4 (whichever model this is) is EXTREMELY efficent with tokens.

Qwen 4 vs Deepseek Flash one shot Christmas Village. Qwen 4 (whichever model this is) is EXTREMELY efficent with tokens.

Loktar 🇺🇸

23,260 次观看 • 5 天前

OpenAI recently released its first open-weights model since GPT-2, entering a field led by DeepSeek and Alibaba's Qwen. Ankit () breaks down these top OSS models, including what sets them apart under the hood: mixture-of-experts, long-context training, and post-training techniques that shape reasoning and alignment—and how different design choices lead to surprisingly similar performance. 00:00 – OpenAI OSS Launch 01:00 – Comparing Open Source LLM Architectures 01:46 – GPT OSS Overview 02:37 – Under The Hood of GPT OSS 03:25 – Qwen-3 Architecture 04:17 – Qwen-3 Training 05:12 – Qwen-3 Post-Training 06:08 – Qwen-3 Reasoning & RL Innovations 06:52 – DeepSeek V3 Overview 07:40 – DeepSeek V3.1 Updates 08:39 – Attention Mechanism (MLA) 09:39 – Comparing Model Sizes 10:35 – Long Context Strategies 11:25 – Reflections on Methods 12:00 – Takeaways

OpenAI recently released its first open-weights model since GPT-2, entering a field led by DeepSeek and Alibaba's Qwen. Ankit () breaks down these top OSS models, including what sets them apart under the hood: mixture-of-experts, long-context training, and post-training techniques that shape reasoning and alignment—and how different design choices lead to surprisingly similar performance. 00:00 – OpenAI OSS Launch 01:00 – Comparing Open Source LLM Architectures 01:46 – GPT OSS Overview 02:37 – Under The Hood of GPT OSS 03:25 – Qwen-3 Architecture 04:17 – Qwen-3 Training 05:12 – Qwen-3 Post-Training 06:08 – Qwen-3 Reasoning & RL Innovations 06:52 – DeepSeek V3 Overview 07:40 – DeepSeek V3.1 Updates 08:39 – Attention Mechanism (MLA) 09:39 – Comparing Model Sizes 10:35 – Long Context Strategies 11:25 – Reflections on Methods 12:00 – Takeaways

Y Combinator

208,865 次观看 • 1 年前

Alright guys, wow. Qwen 3.8 27B is basically Qwen 4. What we have here is one of the best one-shot shark survival games I've made with any model. I ran this same prompt through fable, and the biggest thing all models struggle with in this test is the top-down view of the sharks. Fable even had all of the fins reversed and details jumbled on the first turn. Before now, it was really a 2 to 8 turn test to get something final regardless of model. Qwen 3.8 27B just nailed this in one output, no harness at all, just purely the model thoroughly thinking before outputting the final result. On every asset used, different shark variants, everything, it all looks incredibly tight. The boat movement, shark speed and engagement, everything is so much better than I expected, and I was expecting a lot. Imagine what 8 turns would do in a harness? This confirms to me that while Qwen 3.8 27B thinks a lot, for the first time in a local model on the base weights, IT MIGHT ACTUALLY BE WORTH IT This is a Frontier Lab model result, and it ran on my 5090 at 60tps in Q5_K_M. I am just in complete astoundment at what Qwen has accomplished here, and I am incredibly grateful that this model is open source. The speed will also increase drastically with optimisations. This is just an incredible day. What a blessing. More to come!

Alright guys, wow. Qwen 3.8 27B is basically Qwen 4. What we have here is one of the best one-shot shark survival games I've made with any model. I ran this same prompt through fable, and the biggest thing all models struggle with in this test is the top-down view of the sharks. Fable even had all of the fins reversed and details jumbled on the first turn. Before now, it was really a 2 to 8 turn test to get something final regardless of model. Qwen 3.8 27B just nailed this in one output, no harness at all, just purely the model thoroughly thinking before outputting the final result. On every asset used, different shark variants, everything, it all looks incredibly tight. The boat movement, shark speed and engagement, everything is so much better than I expected, and I was expecting a lot. Imagine what 8 turns would do in a harness? This confirms to me that while Qwen 3.8 27B thinks a lot, for the first time in a local model on the base weights, IT MIGHT ACTUALLY BE WORTH IT This is a Frontier Lab model result, and it ran on my 5090 at 60tps in Q5_K_M. I am just in complete astoundment at what Qwen has accomplished here, and I am incredibly grateful that this model is open source. The speed will also increase drastically with optimisations. This is just an incredible day. What a blessing. More to come!

Kyle Hessling

26,243 次观看 • 2 个月前

Operating System powered by Qwen 3.8 27B at 1950 tokens/sec! here is what 1,950 tokens/second Qwen's 3.8 27b actually looks like on Cerebras: i wrote a minimal python web server that turns cerebras inference into a live operating system. zero apps on disk. when you double click an icon: 1) python proxies a raw SSE stream from qwen 27b at 1,950 tok/s 2) calculator compiles & mounts in 11s 3) full canvas paint studio with brush engine compiles in 10s. at 2,000 tokens/second, software is just an on demand hallucination that runs instantly. the model weights ARE the operating system runtime. what else would you build at 1,950 tokens/second?

Operating System powered by Qwen 3.8 27B at 1950 tokens/sec! here is what 1,950 tokens/second Qwen's 3.8 27b actually looks like on Cerebras: i wrote a minimal python web server that turns cerebras inference into a live operating system. zero apps on disk. when you double click an icon: 1) python proxies a raw SSE stream from qwen 27b at 1,950 tok/s 2) calculator compiles & mounts in 11s 3) full canvas paint studio with brush engine compiles in 10s. at 2,000 tokens/second, software is just an on demand hallucination that runs instantly. the model weights ARE the operating system runtime. what else would you build at 1,950 tokens/second?

Alok

455,156 次观看 • 1 个月前

Open-weight Qwen 3.8 2.4T built a Call of Duty clone in one prompt 🪖 Qwen released the 2.4T Max weights, so we rented a B200 cluster and asked the model to make a Call of Duty clone Output: ~1.1M tokens · 5 hours · one prompt Almost no one can run 2.4T at home, luckily there is a 27B version that runs on a 16GB MacBook Air with our Atomic Dynamic quants

Open-weight Qwen 3.8 2.4T built a Call of Duty clone in one prompt 🪖 Qwen released the 2.4T Max weights, so we rented a B200 cluster and asked the model to make a Call of Duty clone Output: ~1.1M tokens · 5 hours · one prompt Almost no one can run 2.4T at home, luckily there is a 27B version that runs on a 16GB MacBook Air with our Atomic Dynamic quants

atomic.chat

46,359 次观看 • 2 个月前

how i run the local QWEN 3.8 27B via an iPhone the power we have is insane. thanks to china <3

how i run the local QWEN 3.8 27B via an iPhone the power we have is insane. thanks to china <3

Paulius 🏴‍☠️

86,375 次观看 • 2 个月前

MiniCPM-V 4.7 beat Qwen 3.6 35B-3A at Pagoda 🏯 OpenBMB's new 35B-A3 model made a better 3D Voxel Japanese Pagoda while consuming 10% less tokens than Qwen 3.6 35B-3A Run AI models locally ->

MiniCPM-V 4.7 beat Qwen 3.6 35B-3A at Pagoda 🏯 OpenBMB's new 35B-A3 model made a better 3D Voxel Japanese Pagoda while consuming 10% less tokens than Qwen 3.6 35B-3A Run AI models locally ->

atomic.chat

29,612 次观看 • 29 天前

Where the AI model name "Qwen" came from, and here I'm asking why the U.S. doesn't have its own open source version of Qwen. We need something like "Ameriqwen" that was built in America, but we have no large, capable, well-funded open source AI model engineering group in the USA. Instead, we have for-profit corporations that mostly operate as BLACK BOXES that have zero real interest in the public good.

Where the AI model name "Qwen" came from, and here I'm asking why the U.S. doesn't have its own open source version of Qwen. We need something like "Ameriqwen" that was built in America, but we have no large, capable, well-funded open source AI model engineering group in the USA. Instead, we have for-profit corporations that mostly operate as BLACK BOXES that have zero real interest in the public good.

HealthRanger

14,163 次观看 • 1 个月前

When will AI personal assistants be fast enough to be useful? Using Qwen 3.8 27B on Cerebras at ~1,500 tokens/sec, we made an AI personal assistant 19x faster than a suite of other AI personal assistants - Grok Bot, Meta Muse, and Claude Cowork - on the same dinner reservation task.

When will AI personal assistants be fast enough to be useful? Using Qwen 3.8 27B on Cerebras at ~1,500 tokens/sec, we made an AI personal assistant 19x faster than a suite of other AI personal assistants - Grok Bot, Meta Muse, and Claude Cowork - on the same dinner reservation task.

Cerebras

29,430 次观看 • 1 个月前

Qwen 3.8 27B is already agentic lore. So we ran two Qwen 3.8 27B models in one local build. We stacked the Qwen models using the new Windows Hermes Agent desktop bot mode running on our 128 GB Ryzen AI Max+ processor based agent computer. See how Ryzen AI Max+ architecture keeps local Qwen agent builders, building.

Qwen 3.8 27B is already agentic lore. So we ran two Qwen 3.8 27B models in one local build. We stacked the Qwen models using the new Windows Hermes Agent desktop bot mode running on our 128 GB Ryzen AI Max+ processor based agent computer. See how Ryzen AI Max+ architecture keeps local Qwen agent builders, building.

AMD Ryzen

45,176 次观看 • 2 个月前

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.

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

51,685 次观看 • 4 个月前

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses. pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ). That loop is why Qwen quietly became the most downloaded open model family on Hugging Face. Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack. Watch and save it, then read the article below.

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses. pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ). That loop is why Qwen quietly became the most downloaded open model family on Hugging Face. Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack. Watch and save it, then read the article below.

h100envy

114,423 次观看 • 4 个月前

you can just hallucinate the entire internet with Qwen 3.8 27b running at 2,000 tokens/second? part 2 of turning Cerebras + Qwen 3.8 27B into an OS: built an offline browser with zero network calls and mounted it directly the JIT ubuntu desktop. no wifi. no scraping. zero packets sent to external CDNs. you search a site, set a year, and qwen 27b at 1,950 tok/s synthesizes the entire DOM on the fly. here is youtube in 2045 vs 1999: → search google for youtube inside the OS → scrub to 2045: instant futuristic feed → scrub to 1999: raw web 1.0 time capsule in seconds at this speed, browsing isn't retrieving files from a server, it's querying an alternate reality. a 2D browser window is just step zero. imagine full operating systems, virtual worlds, and complex simulation engines existing purely as model weights. zero gigabytes stored on disk, just pure interactive reality streamed on demand. What else becomes a possibility with the qwen 3.8 27b (dense) at 2000 tokens/sec?

you can just hallucinate the entire internet with Qwen 3.8 27b running at 2,000 tokens/second? part 2 of turning Cerebras + Qwen 3.8 27B into an OS: built an offline browser with zero network calls and mounted it directly the JIT ubuntu desktop. no wifi. no scraping. zero packets sent to external CDNs. you search a site, set a year, and qwen 27b at 1,950 tok/s synthesizes the entire DOM on the fly. here is youtube in 2045 vs 1999: → search google for youtube inside the OS → scrub to 2045: instant futuristic feed → scrub to 1999: raw web 1.0 time capsule in seconds at this speed, browsing isn't retrieving files from a server, it's querying an alternate reality. a 2D browser window is just step zero. imagine full operating systems, virtual worlds, and complex simulation engines existing purely as model weights. zero gigabytes stored on disk, just pure interactive reality streamed on demand. What else becomes a possibility with the qwen 3.8 27b (dense) at 2000 tokens/sec?

Alok

1,698,786 次观看 • 1 个月前

The new Qwen 3.5 by Qwen running on-device on iPhone 17 Pro. Qwen 3.5 beats models 4 times its size, has strong visual understanding, and can toggle reasoning on or off. The 2B 6-bit model here is running with MLX optimized for Apple Silicon.

The new Qwen 3.5 by Qwen running on-device on iPhone 17 Pro. Qwen 3.5 beats models 4 times its size, has strong visual understanding, and can toggle reasoning on or off. The 2B 6-bit model here is running with MLX optimized for Apple Silicon.

Adrien Grondin

3,563,233 次观看 • 8 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 • 5 个月前