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。

Mistral OCR 4 just dropped with bounding boxes (their most-requested feature) so I plugged it into my form-filling test as the helper model. Qwen3.6 reasons, Mistral localizes. Result? Boxes detected, fields filled, mostly landing in the lines. Not pixel-perfect. But close? Yeah, I'll call it close.

stevibe

27,309 subscribers

41,473 次观看 • 4 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

Mistral OCR 3 now returns confidence scores, at page level or word level. Know exactly how sure the model is about what it extracted. Available now via API and to test in Mistral Studio.

Mistral OCR 3 now returns confidence scores, at page level or word level. Know exactly how sure the model is about what it extracted. Available now via API and to test in Mistral Studio.

Mistral for Developers

30,200 次观看 • 6 个月前

Introducing Mistral OCR 4:

Introducing Mistral OCR 4:

Mistral for Developers

163,881 次观看 • 4 个月前

I designed a new test specifically for multimodal models: fill out a paper form. And it's much harder than it sounds. This isn't typing into an electronic field that captures your text. The form is just an image. The model has to place each form element: text, checkmarks — at the correct pixel position on the canvas itself. Results: 🟢 Kimi K2.6 → done in 3:45, 16.7k output tokens 🟡 Step 3.7 Flash → half the fields, 57k output tokens 🔴 Gemini 3.5 Flash → 489k output tokens, never finished. I had to kill it. Gemini burned ~29x more output tokens than Kimi on the exact same task, and Kimi's was the only form that actually looked filled out. The test, a mocked application form, contains some challenging parts, such as one-character-per-box fields. I provided every model the same set of tools: > get canvas size > drop probe markers to find coordinates > add text > add checkmarks > move elements > take a screenshot anytime to check their own work > ... etc So it's vision + spatial reasoning + tool use + long context, all at once. Small models (Qwen, Gemma) can't really complete this test, so I skipped them. What happened: > Kimi nailed name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code — placement slightly loose, but content correct. 15 turns. Clean. > Step got maybe half right — fields dropped, "United States" landed in the email line, data floating outside boxes. Burned 1.24M input tokens doing it (81 turns of re-reading the canvas). > Gemini almost got there visually... then spiraled. By turn 40 it was issuing a delete_elements call wiping element IDs 365–425, basically erasing its own work. 31 minutes, 489k output tokens, still streaming. Terminated. The takeaway isn't "Gemini bad." This test is indeed difficult. But token efficiency is capability now. A model that needs 30x the tokens and still can't converge is going to be 30x the cost in production. Kimi K2.6 just quietly did the thing.

I designed a new test specifically for multimodal models: fill out a paper form. And it's much harder than it sounds. This isn't typing into an electronic field that captures your text. The form is just an image. The model has to place each form element: text, checkmarks — at the correct pixel position on the canvas itself. Results: 🟢 Kimi K2.6 → done in 3:45, 16.7k output tokens 🟡 Step 3.7 Flash → half the fields, 57k output tokens 🔴 Gemini 3.5 Flash → 489k output tokens, never finished. I had to kill it. Gemini burned ~29x more output tokens than Kimi on the exact same task, and Kimi's was the only form that actually looked filled out. The test, a mocked application form, contains some challenging parts, such as one-character-per-box fields. I provided every model the same set of tools: > get canvas size > drop probe markers to find coordinates > add text > add checkmarks > move elements > take a screenshot anytime to check their own work > ... etc So it's vision + spatial reasoning + tool use + long context, all at once. Small models (Qwen, Gemma) can't really complete this test, so I skipped them. What happened: > Kimi nailed name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code — placement slightly loose, but content correct. 15 turns. Clean. > Step got maybe half right — fields dropped, "United States" landed in the email line, data floating outside boxes. Burned 1.24M input tokens doing it (81 turns of re-reading the canvas). > Gemini almost got there visually... then spiraled. By turn 40 it was issuing a delete_elements call wiping element IDs 365–425, basically erasing its own work. 31 minutes, 489k output tokens, still streaming. Terminated. The takeaway isn't "Gemini bad." This test is indeed difficult. But token efficiency is capability now. A model that needs 30x the tokens and still can't converge is going to be 30x the cost in production. Kimi K2.6 just quietly did the thing.

stevibe

25,494 次观看 • 5 个月前

I needed an app to export markdown from images and PDFs, so I built one using the best model on the market: Mistral OCR. It can convert images, PDFs, and even extract tables and images, keeping them contextually within the markdown. All you need is a Mistral AI API.

I needed an app to export markdown from images and PDFs, so I built one using the best model on the market: Mistral OCR. It can convert images, PDFs, and even extract tables and images, keeping them contextually within the markdown. All you need is a Mistral AI API.

Pietro Schirano

48,877 次观看 • 1 年前

Most people think a 0.8B model is a toy. 3rd test. This time: OCR I fed Qwen 3.5 0.8B three images: a Kindle page, my terrible handwriting x 2. Mostly correct. Running locally. From a model smaller than your browser tab.

Most people think a 0.8B model is a toy. 3rd test. This time: OCR I fed Qwen 3.5 0.8B three images: a Kindle page, my terrible handwriting x 2. Mostly correct. Running locally. From a model smaller than your browser tab.

stevibe

125,358 次观看 • 7 个月前

Qwen3.6 35B-A3B dropped yesterday, so I ran it on 4 GPUs to see how it performs: 🟣 RTX 3090 — 49.78 tok/s, TTFT 852ms 🟡 RTX 4090 — 118.93 tok/s, TTFT 686ms 🟢 RTX 5090 — 160.37 tok/s, TTFT 409ms 🔵 DGX Spark — 59.98 tok/s, TTFT 228ms I went with ollama as the backend because honestly, it's the easiest way for most people to get started. One command, model pulled, done. I used Q4_K_M (24GB) across all four cards. The reason is the 3090 and 4090 don't support NVFP4 (only the 5090 and DGX Spark could use it). Keeping the same quant everywhere felt like the fairest way to compare. And yes, you can absolutely squeeze more performance out of every card with vLLM, SGLang, or TensorRT-LLM. But that's not what this test is about. This is just the out-of-the-box experience for folks who own a GPU and want to try the new model tonight.

Qwen3.6 35B-A3B dropped yesterday, so I ran it on 4 GPUs to see how it performs: 🟣 RTX 3090 — 49.78 tok/s, TTFT 852ms 🟡 RTX 4090 — 118.93 tok/s, TTFT 686ms 🟢 RTX 5090 — 160.37 tok/s, TTFT 409ms 🔵 DGX Spark — 59.98 tok/s, TTFT 228ms I went with ollama as the backend because honestly, it's the easiest way for most people to get started. One command, model pulled, done. I used Q4_K_M (24GB) across all four cards. The reason is the 3090 and 4090 don't support NVFP4 (only the 5090 and DGX Spark could use it). Keeping the same quant everywhere felt like the fairest way to compare. And yes, you can absolutely squeeze more performance out of every card with vLLM, SGLang, or TensorRT-LLM. But that's not what this test is about. This is just the out-of-the-box experience for folks who own a GPU and want to try the new model tonight.

stevibe

404,026 次观看 • 6 个月前

The best way to bring the composition from your head into an image → Ideogram V4 + drawing bounding boxes in Comfy. The control here is quite unique. The model uses structured JSON so drawing bounding boxes to get the exact placement works very well. The model only needs 12 steps (turbo), so iterating with different seeds + very impressive text rendering capability leads me to say this is a state of the art open source image model right now. Using the 'Ideogram 4 Prompt Builder' node by Kijai.

The best way to bring the composition from your head into an image → Ideogram V4 + drawing bounding boxes in Comfy. The control here is quite unique. The model uses structured JSON so drawing bounding boxes to get the exact placement works very well. The model only needs 12 steps (turbo), so iterating with different seeds + very impressive text rendering capability leads me to say this is a state of the art open source image model right now. Using the 'Ideogram 4 Prompt Builder' node by Kijai.

rob - comfyui

25,841 次观看 • 4 个月前

Fable 5 just made the best Remotion video I have ever generated with ANY model. It is not close. It cloned BridgeSpace pixel perfect. It nailed every selling point. It designed the motion, the pacing, the whole thing. The design taste and creativity of this model are simply unmatched. Watch this and remember: an AI made this.

Fable 5 just made the best Remotion video I have ever generated with ANY model. It is not close. It cloned BridgeSpace pixel perfect. It nailed every selling point. It designed the motion, the pacing, the whole thing. The design taste and creativity of this model are simply unmatched. Watch this and remember: an AI made this.

BridgeMind

81,063 次观看 • 3 个月前

What does it take to bring open models into the enterprise? On the NVIDIA AI Podcast, Mistral CTO and co-founder Timothée Lacroix joins to discuss Mistral AI’s open-model philosophy, its Forge customization framework, and its collaboration with NVIDIA through the Nemotron Coalition.

What does it take to bring open models into the enterprise? On the NVIDIA AI Podcast, Mistral CTO and co-founder Timothée Lacroix joins to discuss Mistral AI’s open-model philosophy, its Forge customization framework, and its collaboration with NVIDIA through the Nemotron Coalition.

NVIDIA

32,297 次观看 • 4 个月前

The #SPX6900 thesis is playing out in real time: “The perfect meme token.” “It just checks off so many boxes.” /////////////////////////////////////////////////////// Stop trading and believe in something. @SPX6900

The #SPX6900 thesis is playing out in real time: “The perfect meme token.” “It just checks off so many boxes.” /////////////////////////////////////////////////////// Stop trading and believe in something. @SPX6900

0xpepethefrog 💹🧲

101,270 次观看 • 2 年前

Mistral just shipped Large 4, its biggest model ever. So I gave the same 3D prompt to 8 models. One shot, no fixes. Open: Mistral Large 4, Kimi K3, GLM 5.3, DeepSeek V4 Pro Closed: Claude Opus 5.5, Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol Results below 👇

Mistral just shipped Large 4, its biggest model ever. So I gave the same 3D prompt to 8 models. One shot, no fixes. Open: Mistral Large 4, Kimi K3, GLM 5.3, DeepSeek V4 Pro Closed: Claude Opus 5.5, Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol Results below 👇

Dominique Silvestre

45,010 次观看 • 20 天前

StockX just launched Mystery Boxes, so I bought the most expensive one. Here’s what happened:

StockX just launched Mystery Boxes, so I bought the most expensive one. Here’s what happened:

Kari

32,116 次观看 • 3 个月前

New body-physics test for the best AI video tools: The newly released Grok Imagine 1.5, the king Seedance 2.0, Google’s “revolutionary super-duper” Gemini Omni Flash, and the old-timer Kling 3.0 Pro. This time I tested how each model handles realistic running motion, body movement, fabric physics, and natural secondary motion. And the result was much less obvious than I expected. Believe it or not, I can’t call Seedance the clear winner this time. Each model got several attempts. - Grok Imagine 1.5: The new version is finally available on the official Grok website, so I could properly test it. Honestly, I didn’t notice a massive improvement over the previous version. Still, the result was acceptable. It produced the most cartoonish image, and the woman runs as if she’s wearing heels, but the body physics were decent enough. It also understood the instructions quickly and followed them correctly. - Kling 3.0 Pro: The old man decided to test my patience. It repeatedly blocked a simple running scene as adult content, then misunderstood the instructions several times. The successful result has the most realistic lighting and frame rate, but the actual body physics look strange. It almost feels like loose foam padding is bouncing inside the leggings. There are also several visible artifacts and unnatural movements. - Gemini Omni Flash: As usual, it gave me that strange slow-motion, low-FPS look that Google models seem to love. But it didn’t censor anything, understood the instructions immediately, and produced a beautiful, realistic result. Surprisingly, this is the output I liked the most in this test. - Seedance 2.0: Seedance also blocked a couple of generations, just like Kling, but eventually produced a strong result. It delivered the most beautiful and visually appealing footage, but I honestly expected better physics. The video looks great, yet I can’t confidently call it the most physically accurate result. - My ranking: 1. Gemini Omni Flash — not perfect, but the best overall result for me 2. Seedance 2.0 — visually stunning, but I preferred Omni’s physics 3. Grok Imagine 1.5 — cartoonish look and strange running, but still acceptable 4. Kling 3.0 Pro — the longest wait, the strangest physics, and the most inconsistent result Do you agree with my ranking? #AIVideo

New body-physics test for the best AI video tools: The newly released Grok Imagine 1.5, the king Seedance 2.0, Google’s “revolutionary super-duper” Gemini Omni Flash, and the old-timer Kling 3.0 Pro. This time I tested how each model handles realistic running motion, body movement, fabric physics, and natural secondary motion. And the result was much less obvious than I expected. Believe it or not, I can’t call Seedance the clear winner this time. Each model got several attempts. - Grok Imagine 1.5: The new version is finally available on the official Grok website, so I could properly test it. Honestly, I didn’t notice a massive improvement over the previous version. Still, the result was acceptable. It produced the most cartoonish image, and the woman runs as if she’s wearing heels, but the body physics were decent enough. It also understood the instructions quickly and followed them correctly. - Kling 3.0 Pro: The old man decided to test my patience. It repeatedly blocked a simple running scene as adult content, then misunderstood the instructions several times. The successful result has the most realistic lighting and frame rate, but the actual body physics look strange. It almost feels like loose foam padding is bouncing inside the leggings. There are also several visible artifacts and unnatural movements. - Gemini Omni Flash: As usual, it gave me that strange slow-motion, low-FPS look that Google models seem to love. But it didn’t censor anything, understood the instructions immediately, and produced a beautiful, realistic result. Surprisingly, this is the output I liked the most in this test. - Seedance 2.0: Seedance also blocked a couple of generations, just like Kling, but eventually produced a strong result. It delivered the most beautiful and visually appealing footage, but I honestly expected better physics. The video looks great, yet I can’t confidently call it the most physically accurate result. - My ranking: 1. Gemini Omni Flash — not perfect, but the best overall result for me 2. Seedance 2.0 — visually stunning, but I preferred Omni’s physics 3. Grok Imagine 1.5 — cartoonish look and strange running, but still acceptable 4. Kling 3.0 Pro — the longest wait, the strangest physics, and the most inconsistent result Do you agree with my ranking? #AIVideo

Alpha Mom

159,187 次观看 • 4 个月前

We used Mistral OCR with Claude 3.7 to create blog-style overviews for arXiv papers Generate beautiful research blogs with figures, key insights, and clear explanations from the paper with just one click Understand papers in minutes - not hours

We used Mistral OCR with Claude 3.7 to create blog-style overviews for arXiv papers Generate beautiful research blogs with figures, key insights, and clear explanations from the paper with just one click Understand papers in minutes - not hours

alphaXiv

153,847 次观看 • 1 年前