Video yükleniyor...

Video Yüklenemedi

Bu video yüklenirken bir sorun oluştu. Bu geçici bir ağ sorunundan kaynaklanıyor olabilir veya video kullanılamıyor olabilir.

Ana Sayfaya Dön

Still didn't attempt any quality testing, but let's say that GPT 5.6 Sol was able almost unassisted to write this Laguna S2.1 inference implementation following the other two models in DwarfStar, and it is very fast, 50 t/s generation, very fast prefill as well.

antirez

81,933 subscribers

43,676 görüntüleme • 2 ay önce •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

22 Yorum

Sakura Yuki profil fotoğrafı

Sakura Yuki2 ay önce

50 t/s is nice, but Sol writing Laguna support from two neighboring implementations is the wilder result. Inference backends are becoming executable architecture docs.

Giovanni profil fotoğrafı

Giovanni2 ay önce

I hope your version will change my mind, but for my use case, Laguna has been a disaster so far!

antirez profil fotoğrafı

antirez2 ay önce

I'm also seeing odd things for the Q4 quants.

Giovanni profil fotoğrafı

Giovanni2 ay önce

Same experience with Q4.

Shubham Arora profil fotoğrafı

Shubham Arora2 ay önce

That’s as fast as gpt-5.6-luna api !

Michał Piszczek profil fotoğrafı

Michał Piszczek2 ay önce

50 t/s is a marketing number until the quality pass happens. fast inference on wrong output is just a quicker way to be wrong.

BitXChange profil fotoğrafı

BitXChange2 ay önce

Sol is extremely talented and helpful goblin. Substantially less mistakes and bs than Fable.

inside inside profil fotoğrafı

inside inside2 ay önce

When are you going to implement --n-cpu-moe ? Can GPT 5.6-Sol implement it ?

Ɠuglièlmo "𝚐𝚠𝚣" Ɲígri profil fotoğrafı

Ɠuglièlmo "𝚐𝚠𝚣" Ɲígri2 ay önce

Impressive work! I understand this is Q4: how does it compare to the other local models?

Francesco Marasco 🇮🇹🇪🇺 profil fotoğrafı

Francesco Marasco 🇮🇹🇪🇺2 ay önce

Have you had a chance to benchmark the models on DwarfStar for both speed and reasoning?

LeetLLM.com profil fotoğrafı

LeetLLM.com2 ay önce

50 t/s is nice, but the real eval is the cleanup diff. How much of Sol’s implementation survived your review unchanged?

clainstone profil fotoğrafı

clainstone2 ay önce

Do you think it's still useful to learn how to write complex kernels?

pls seed profil fotoğrafı

pls seed2 ay önce

Can you run the tests ? On mine it's worse then flash V4.

Diego Carlino profil fotoğrafı

Diego Carlino2 ay önce

what hardware ?

antirez profil fotoğrafı

antirez2 ay önce

m5 max 128gb

Angel profil fotoğrafı

Angel2 ay önce

united open agents

Dustin Ogle profil fotoğrafı

Dustin Ogle2 ay önce

Heard mixed things about the quality. Coding is great. Other stuff depends.

Ljubomir Josifovski profil fotoğrafı

Ljubomir Josifovski2 ay önce

how is the decline with context depth increase - in absolute, and compared to ds4f?

Daksh Trehan profil fotoğrafı

Daksh Trehan2 ay önce

correctness in inference code was never the hard part. models have written plausible attention loops for a year now. 50 t/s with fast prefill on an m5 max is a different animal. kv cache layout, the 4-bit paths, kernels that don't stall on metal. the stuff you only get right after sitting with a profiler for a week. if sol pulled that off almost unprompted, then the last skill i figured still needed a human is going too. hardware-aware perf work. the unglamorous profiling nobody wants to touch.

shaped profil fotoğrafı

shaped2 ay önce

sol cooked up its own rocket engine

noname profil fotoğrafı

noname2 ay önce

Give a shot to benchlocal-cli

Jedd Haberstro profil fotoğrafı

Jedd Haberstro2 ay önce

Would love to hear about how you go about prompting these days.

Benzer Videolar

I didn't expect DeepSeek v4 PRO (not Flash) to run well on the Mac Studio M3 Ultra with 512GB of RAM. This is 2 bit quantized with the same DwarfStar recipe used for Flash. 433GB GGUF file. 130 t/s prefill, 13 t/s generation. Prefill in the video is low because small prompt.

I didn't expect DeepSeek v4 PRO (not Flash) to run well on the Mac Studio M3 Ultra with 512GB of RAM. This is 2 bit quantized with the same DwarfStar recipe used for Flash. 433GB GGUF file. 130 t/s prefill, 13 t/s generation. Prefill in the video is low because small prompt.

antirez

172,816 görüntüleme • 4 ay önce

For the DGX Spark owners. This is what you get with DS4 in your hardware. I want to post this to show how with fast prefill and not very fast generation, the system remains absolutely fine to use.

For the DGX Spark owners. This is what you get with DS4 in your hardware. I want to post this to show how with fast prefill and not very fast generation, the system remains absolutely fine to use.

antirez

40,615 görüntüleme • 4 ay önce

well, i connected GPT-5.6 Sol to Blender and... let's just say the world of 3D graphics (at least via Blender's MCP) is safe for now... but still... this whole scene was generated by GPT-5.6 with very little human intervention feels like *EARLY* pixar, like too early lol

well, i connected GPT-5.6 Sol to Blender and... let's just say the world of 3D graphics (at least via Blender's MCP) is safe for now... but still... this whole scene was generated by GPT-5.6 with very little human intervention feels like EARLY pixar, like too early lol

Gavin Purcell

71,458 görüntüleme • 2 ay önce

DeepSeek v4 PRO running in DwarfStar on a DGX Station. Can go faster, already peaks 50 t/s without DSpark: dynamic VRAM/RAM allocation of experts based on past statistics, 4k t/s prefill on big chunks. VRAM/RAM layer-major streaming. Mixed RAM/VRAM inference requires care.

DeepSeek v4 PRO running in DwarfStar on a DGX Station. Can go faster, already peaks 50 t/s without DSpark: dynamic VRAM/RAM allocation of experts based on past statistics, 4k t/s prefill on big chunks. VRAM/RAM layer-major streaming. Mixed RAM/VRAM inference requires care.

antirez

26,763 görüntüleme • 1 ay önce

meta just dropped a bomb. Muse Spark 1.1 is 86% cheaper than GPT-5.6 Sol and 91% cheaper than Fable 5 on output tokens ($4.25 vs $30 vs $50 per M) I thought the model was going to be horrendous, but it ended up being insanely fast, and very capable. front end design is VERY strong. Agentic capabilities seem great as it was able to spawn and even orchestrate other agents inside my app cnvs using the mcp and cli tool with zero instruction. we have soo many insanely good models right now.

meta just dropped a bomb. Muse Spark 1.1 is 86% cheaper than GPT-5.6 Sol and 91% cheaper than Fable 5 on output tokens ($4.25 vs $30 vs $50 per M) I thought the model was going to be horrendous, but it ended up being insanely fast, and very capable. front end design is VERY strong. Agentic capabilities seem great as it was able to spawn and even orchestrate other agents inside my app cnvs using the mcp and cli tool with zero instruction. we have soo many insanely good models right now.

Max Blade

88,802 görüntüleme • 2 ay önce

DGX Station: this is just the start of what it is possible to do (and why I claim DwarfStar could be *the* Station inference engine). DeepSeek v4 PRO Q2 with routed experts split among VRAM / RAM with kernels optimized for this peculiar setup. 45 t/s but can go faster.

DGX Station: this is just the start of what it is possible to do (and why I claim DwarfStar could be the Station inference engine). DeepSeek v4 PRO Q2 with routed experts split among VRAM / RAM with kernels optimized for this peculiar setup. 45 t/s but can go faster.

antirez

41,545 görüntüleme • 1 ay önce

Crazy prefill speed of DwarfStar running DeepSeek v4 PRO in the DGX Station (private branch for now) using techniques that leverage the peculiar RAM + VRAM mixed setup of the Station. 3000 t/s but even more can be reached.

Crazy prefill speed of DwarfStar running DeepSeek v4 PRO in the DGX Station (private branch for now) using techniques that leverage the peculiar RAM + VRAM mixed setup of the Station. 3000 t/s but even more can be reached.

antirez

63,235 görüntüleme • 1 ay önce

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

antirez

84,127 görüntüleme • 5 ay önce

Sam Altman says: "GPT-5.5 was maybe as good as an average math professor, 5.6 as good as a top 1–2% one — and the internal model past Astra can do things even the best mathematicians cannot" Math has seen an unusually fast takeoff, but by almost any definition, this is a fast takeoff

Sam Altman says: "GPT-5.5 was maybe as good as an average math professor, 5.6 as good as a top 1–2% one — and the internal model past Astra can do things even the best mathematicians cannot" Math has seen an unusually fast takeoff, but by almost any definition, this is a fast takeoff

Haider.

44,153 görüntüleme • 21 gün önce

Two frontier models. Two very different demos. (GPT-5.6 Sol and Claude Fable 5) GPT-5.6 Sol Ultra was pushed in a mathematical direction and generated a Minecraft-style clone in Lean. Yes, that Lean. Meanwhile, Claude Fable 5 took a game development route and built a Minecraft mod featuring a Dark Souls-inspired enemy with multiple attack patterns, combos, and combat logic. They're solving completely different problems, but both are great examples of how capable today's models have become. The pace of progress is remarkable.

Two frontier models. Two very different demos. (GPT-5.6 Sol and Claude Fable 5) GPT-5.6 Sol Ultra was pushed in a mathematical direction and generated a Minecraft-style clone in Lean. Yes, that Lean. Meanwhile, Claude Fable 5 took a game development route and built a Minecraft mod featuring a Dark Souls-inspired enemy with multiple attack patterns, combos, and combat logic. They're solving completely different problems, but both are great examples of how capable today's models have become. The pace of progress is remarkable.

FHILY👑

19,336 görüntüleme • 2 ay önce

First kinda working implementation of GLM 5.2 in DwarfStar. Will take some time to be good enough, but it is a promising start. 433 GB GGUF file. M3 Ultra 512GB.

First kinda working implementation of GLM 5.2 in DwarfStar. Will take some time to be good enough, but it is a promising start. 433 GB GGUF file. M3 Ultra 512GB.

antirez

69,587 görüntüleme • 3 ay önce

GPT-5.6 vs GPT-5.5 on my custom spaceship prompt. I gave both models the exact same custom prompt. This is also the same prompt I previously gave to Fable 5. For context, GPT-5.6 Pro worked for 87 minutes, while GPT-5.5 Extra High worked for 34 minutes and 42 seconds. As I’ve said before, based on great authority GPT-5.6 will be an incremental/soldi improvement over GPT-5.5, not a “Fable killer.” My rough expectation has been that it would trade blows with Fable 5 on some benchmarks, maybe win around half depending on the category, but not clearly surpass it overall. And again fable five will have bigger model smell, but this was expected. After testing this coding output, that view feels pretty accurate. GPT-5.6 is clearly better than GPT-5.5 in several visual areas. The lighting, shading, chairs, object details, and exterior of the spaceship looked noticeably stronger. The scene was also easier to test. I do want to give GPT-5.5 credit though. It built out the rooms much much better and the planets looked better than GPT-5.6’s. It was also interesting that both GPT-5.5 and GPT-5.6 produced better-looking planets than Fable 5 in this specific test. The downside with GPT-5.5 was stability. The game was much glitchier and harder to test compared to GPT-5.6. But when it comes to the core of the demo, which is the spaceship itself, Fable 5 still beat both models pretty comfortably. GPT-5.6 is impressive, but from this test, it looks exactly like what I expected which was a meaningful incremental improvement over GPT-5.5, at least for indie game demos, but not something that replaces Fable 5. In collaboration with Chetaslua

GPT-5.6 vs GPT-5.5 on my custom spaceship prompt. I gave both models the exact same custom prompt. This is also the same prompt I previously gave to Fable 5. For context, GPT-5.6 Pro worked for 87 minutes, while GPT-5.5 Extra High worked for 34 minutes and 42 seconds. As I’ve said before, based on great authority GPT-5.6 will be an incremental/soldi improvement over GPT-5.5, not a “Fable killer.” My rough expectation has been that it would trade blows with Fable 5 on some benchmarks, maybe win around half depending on the category, but not clearly surpass it overall. And again fable five will have bigger model smell, but this was expected. After testing this coding output, that view feels pretty accurate. GPT-5.6 is clearly better than GPT-5.5 in several visual areas. The lighting, shading, chairs, object details, and exterior of the spaceship looked noticeably stronger. The scene was also easier to test. I do want to give GPT-5.5 credit though. It built out the rooms much much better and the planets looked better than GPT-5.6’s. It was also interesting that both GPT-5.5 and GPT-5.6 produced better-looking planets than Fable 5 in this specific test. The downside with GPT-5.5 was stability. The game was much glitchier and harder to test compared to GPT-5.6. But when it comes to the core of the demo, which is the spaceship itself, Fable 5 still beat both models pretty comfortably. GPT-5.6 is impressive, but from this test, it looks exactly like what I expected which was a meaningful incremental improvement over GPT-5.5, at least for indie game demos, but not something that replaces Fable 5. In collaboration with Chetaslua

Chris

250,919 görüntüleme • 3 ay önce

Took the good work of the communtiy of DwarfStar and consolidating the Strix Halo support. It looks very good. More QA in the next days and the final merge soon.

Took the good work of the communtiy of DwarfStar and consolidating the Strix Halo support. It looks very good. More QA in the next days and the final merge soon.

antirez

17,567 görüntüleme • 4 ay önce

Nvidia $NVDA CEO Jensen Huang was asked about a recent story that said he warned that China will beat the US in the AI race “That’s not what I said. What I said was China has very good AI technology. They have many AI researchers, in fact 50% of the world’s AI researchers are in China. And they develop very good AI technology. In fact, the most popular AI models in the world today, open-source models, are from China. So, they are moving very, very fast. The United States has to continue to move incredibly fast. And otherwise, otherwise – the world is very competitive, so we have to run fast.” (H/T Dan Nystedt)

Nvidia $NVDA CEO Jensen Huang was asked about a recent story that said he warned that China will beat the US in the AI race “That’s not what I said. What I said was China has very good AI technology. They have many AI researchers, in fact 50% of the world’s AI researchers are in China. And they develop very good AI technology. In fact, the most popular AI models in the world today, open-source models, are from China. So, they are moving very, very fast. The United States has to continue to move incredibly fast. And otherwise, otherwise – the world is very competitive, so we have to run fast.” (H/T Dan Nystedt)

Evan

1,946,992 görüntüleme • 11 ay önce

GPT-6 makes me scared as a 3D artist, cuz it's freaking insane. I mean, I'm testing this first hand everyday, but this time it is massive leap. look, 4 month ago I was testing GPT & Claude in Blender and they were not able to position simple objects in simple scene.. And look now... Whole car assembled from primitives in ONE prompt and animate it (hate to say that, but it was indeed one prompt). Yes, yes, it is not perfect and maybe not really usable still at this stage. BUT WHAT A JUMP.

GPT-6 makes me scared as a 3D artist, cuz it's freaking insane. I mean, I'm testing this first hand everyday, but this time it is massive leap. look, 4 month ago I was testing GPT & Claude in Blender and they were not able to position simple objects in simple scene.. And look now... Whole car assembled from primitives in ONE prompt and animate it (hate to say that, but it was indeed one prompt). Yes, yes, it is not perfect and maybe not really usable still at this stage. BUT WHAT A JUMP.

Stefan 3D AI

189,058 görüntüleme • 1 ay önce

DwarfStar running DeepSeek v4.1 Flash on a 128GB M5 Max. I didn't expect with SSD streaming it could be so fast. Recent SSD streaming changes to retain the right experts surely helped, but also maybe DS4.1 uses the same experts more. Will push online when ready QA > ASAP.

DwarfStar running DeepSeek v4.1 Flash on a 128GB M5 Max. I didn't expect with SSD streaming it could be so fast. Recent SSD streaming changes to retain the right experts surely helped, but also maybe DS4.1 uses the same experts more. Will push online when ready QA > ASAP.

antirez

432,170 görüntüleme • 27 gün önce

Interview about the March 31: "It was really fast. No one was ready for it. Beck has a reflex instinct. I didn't mean to do anything to them, but just like, "Who should not be here?" or something like that. Very shocked, I was very shocked because I felt that it was wrong. You shouldn't do this and you shouldn't do this to any artist. It's something like that. The people who are afraid are afraid. It's not okay. We live like a lovely fan club, we love each other very much and trust each other very much. We are like a family. I don't want anyone to do this to us" BECKY IS WORTH IT #LOrealParisTHxBecky #beckysangels

Interview about the March 31: "It was really fast. No one was ready for it. Beck has a reflex instinct. I didn't mean to do anything to them, but just like, "Who should not be here?" or something like that. Very shocked, I was very shocked because I felt that it was wrong. You shouldn't do this and you shouldn't do this to any artist. It's something like that. The people who are afraid are afraid. It's not okay. We live like a lovely fan club, we love each other very much and trust each other very much. We are like a family. I don't want anyone to do this to us" BECKY IS WORTH IT #LOrealParisTHxBecky #beckysangels

D_ 🐰🦦

43,805 görüntüleme • 1 yıl önce

I'm blown away by Qwen 3.8 27B. A few months ago when I was testing open-weight models, it seemed clear that only GLM and Kimi class models (+1T parameters) could get close to GPT-5.6 / Opus / Fable levels. And now this tiny 27B model gets almost as close. I gotta say, China got this down to an art. Not just benchmarks. I actually tested this in my companies 👇

I'm blown away by Qwen 3.8 27B. A few months ago when I was testing open-weight models, it seemed clear that only GLM and Kimi class models (+1T parameters) could get close to GPT-5.6 / Opus / Fable levels. And now this tiny 27B model gets almost as close. I gotta say, China got this down to an art. Not just benchmarks. I actually tested this in my companies 👇

Simon Høiberg

16,325 görüntüleme • 1 ay önce

GPT-6.1 Sol isn’t that good I asked it to build the Eiffel Tower in Three.js, and this is what it came up with. It performs pretty well on some other tests, but it failed badly on this scene compared to other models. For me, it still feels below Astra and even Kimi in frontend and 3D. OpenAI says it’s more efficient, but it seems to consume a lot of tokens.

GPT-6.1 Sol isn’t that good I asked it to build the Eiffel Tower in Three.js, and this is what it came up with. It performs pretty well on some other tests, but it failed badly on this scene compared to other models. For me, it still feels below Astra and even Kimi in frontend and 3D. OpenAI says it’s more efficient, but it seems to consume a lot of tokens.

Bhavy☄️

38,564 görüntüleme • 8 gün önce