
AJ
@ItsmeAjayKV • 11,171 subscribers
Bullish on local AI Independent AI R&D | AI NEWS quants, repe, kernel. 1x 3090: 64GB RAM https://t.co/btCRUhd2GB https://t.co/6sXjIkjujn N4 (日本語勉強中) 🎌
Videos

Hy4-Flash vs DeepSeek-v4.1-Flash-1017 ! Left: deepseek-v4.1-flash-1017 (high) Right: hy4-flash (high) Both one-shot, no harness, no tool calls, same prompt to create a mechanical flying dragonfly ! Two flash models, but way different token count for same task. Looking at the result itself, deepseek-v4.1-flash one is better, it's more mechanical. Hy4-flash does the job, but it's more basic and fall short on the expectation. When it comes to token usage: Deepseek-v4.1-flash: think 68,304 | answer 14,708 | out 83,012 Hy4-flash: think 15,456 | answer 5,205 | out 20,661
AJ12,806 görüntüleme • 1 gün önce

Here's a comparison i never thought i would make Qwen-3.8-27b vs Gemini-3.7-flash Yes, you read it right! I don't know what to say, this is soo damn impressive. the difference is hugeee🤯 Take a look at this result 😍 (watch it in highest quality) Same prompt, both one shot on web-chat ui Qwen-3.8-27b produces consistent beautiful, visually stunning results. This is one of the best water surface simulations i have seen across every local models i have run and at the same time beating Gemini-3.7-flash result. And this is running on my 3090. And yes its Q5, not even f8.
AJ304,015 görüntüleme • 3 ay önce

Qwen4-max-preview's one-shot capabilities are looking INSANE 🔥!! This is just my first impression, but i'm already impressed by what Qwen has done here. This is the best result i have gotten for this lighthouse island prompt out of all the models i have tested. I never asked for this much controls in prompt, but it did, and all of them works, a richer scene altogether. As all my one-shot tests goes, this is made without any harness, direct api call, high thinking effort, no external assets used. This is also way better than DeepSeek-v4.1-flash-1017, it should be considering this is really big. What stands out even more is how qwen4-max-preview uses way less tokens compared to deepseek-flash.
AJ17,511 görüntüleme • 7 gün önce

DeepSeek is officially good on my 3090 ! Running on my 24GB RTX 3090 + 64GB DDR5 RAM ! This is one-shot by deepseek-v4.1-flash Maya-S quant, no harness. With Project-Maya as backend. Shoutout to Peasant Smith and Need More VRAM ! Crazy part is this is with low thinking effort. I'm really happy with the result, didn't think this would run on my 3090 or i'll get this good result. This is just preview, i'm doing more testing.
AJ27,099 görüntüleme • 28 gün önce

Did DeepSeek-v4.1-flash get updated again? I know it's 3D capabilities are good with latest 1017, but not this good. Insane single shot Pokemon Pallet town generation. Simple prompt, asking it to generate a 3d construction of Pallet town, all procedural, without using any external assets. Map is wrong, but 3D build is crazy. Considering this is generated without harness and no do-overs and without any tool calling this is crazy. -Direct deepseek api call, model id "deepseek-flash" Thinking effort: max Total tokens: 140k Total time: 10 min
AJ17,522 görüntüleme • 17 gün önce

WTH Looks like MiniMax (official) cooked with this release. MiniMax-M3.1-Flash just build be the most visually stunning Pagoda out of all other models i have tested. This result is better than DeepSeek-v4.1-flash It's also one-shot, without any extra turns or tool calls, all happened without any harness as well. Direct api call, max thinking, 50K tokens total
AJ28,138 görüntüleme • 1 ay önce

DeepSeek Deepseek-v4.1-flash-v2 coming 👀!! I'm more convinced something has changed behind the deepseek-platform api. The "deepseek-flash" endpoint seems to be routing to a different model now. It could be: - Deepseek-v4.2-flash - An updated checkpoint of v4.1-flash Since it's also a little slower, i wonder if V4.1-pro is being quietly tested behind the endpoint. But either way, i'm getting way better result, the difference is noticeable. See the comparison video, same prompt, same settings.
AJ22,059 görüntüleme • 1 ay önce

WTF DeepSeek cooked really well with new flash update. DeepSeek-v4.1-flash-1017 in 3d game/world building is fucking amazing !! It's so close to Opus-5.5, and beats every other best open model in 3D game development. The textures look like they are imported, but none of these have any external assets at all, all build in code. This is also not using Blender or other softwares. What a upgrade !! See this fps game build with dsh as harness.
AJ16,948 görüntüleme • 1 ay önce

I'm really really impressed with Qwen3.8-Flash-Next😍 It is better than Qwen3.8-27B !! Below is a screen-recording from a Three.js FPS game made by Qwen3.8-Flash-Next running on my 3090. Everything here is generated at runtime: textures, sounds, music, models and the rigid-body solver. No image, audio or mesh files are downloaded. three.js is loaded from a CDN. atomic.chat Qwen3.8-Flash-Next IQ4_XS👌 DeepSeek Harness 👌 Single RTX 3090 + 64GB RAM (Total 88 GB) Stats: - ~ 10 - 11 hrs in total - 22M tokens, 94% cache - Avg decode 20t/s - Avg prefill 200t/s - TTFT avg 9.6s It is slow, but faster than Unsloth IQ3_XXS !
AJ18,718 görüntüleme • 2 ay önce

🤯Laguna S-2.1 UD-IQ3_S on a single 3090 ~30-35t/s decode ~280-290t/s prefill ~23GB VRAM @ 64k context, f16 KV llama.cpp: -ngl 999 + --n-cpu-moe 32 + fa on 118B MoE, ~8B active, did we get a decent mid-range model that doesn't need multi gpu and can run with good enough speeds for long-context coding tasks? Might actually stick with this one. Been waiting for a mid-range model that fits on a single 3090, this is it. Usual three.js tests underway. Update soon.
AJ20,339 görüntüleme • 3 ay önce

For the first time, i'm not even bothered about missing Fable 5.1 Because i now have Qwen Qwen3.8-Flash-Next with me🤯! PS: Single 3090 users, you might not want to skip this one, you're in for a treat ! Gap between frontier closed models and local models is getting really really small. Building a Rocket League type game in Three.js 🚗⚽️ Gave it a starting prompt and let agent cook. It built an entire Rocket League style experience from scratch. 3D arena, car physics, ball physics, boost system, double-jump flips, goals, kickoff screen, AI apponent, 3- min matches, overtime, score hud, multiple camera modes, particles and sfx. You know the crazy part? This is Qwen3.8-Flash-Next running locally on my single 3090. You know what quant i'm using ? It's Unsloth AI UD-IQ3_XXS💀 7.27M tokens, ~6.89 cached in 2h 30min. Local AI is getting ridiculous🚀
AJ12,914 görüntüleme • 2 ay önce

I'm really happy with Qwen3.8-27B ! All frustrations of not able to use Ox-Alpha and running out of credits for Glm-5.3, Grok turned me into Qwen-3.8-27b. Running locally on my 3090. This is the first fps game i'm creating with Qwen3.8-27b. I was after NMS vibe, alien planet, vast, flying drones, background music + sound effects. Without any external assets at all. With proper game like graphics settings panel 😍 And i'm really impressed, because all this is generated with my mixed quant for Qwen3.8-27B, which gives me more context + speed for a really good quality work. DSK harness at beginning and Pi agent towards the end cause i wanted to get this done fast. Full quant name: HF aj9o9/Qwen3.8-27B-GGUF file name: Qwen3.8-27B-gdn8-q6attn-iq3ffn.gguf Just like all of you, i also hate having to compact context, so higher context without ever needing to compromise quality is what i'm always after and this quant is created for that. Gives me 210K context at q8_0 kv + dflash2 Q4_K_M for speculative decoding. This result in 23GB vram usage.
AJ12,784 görüntüleme • 2 ay önce

Y’all thought I stopped at Q4_K_M 😭 Nope, q3 and q4 didn't cut it, quality was not at all good on my three js tests. So now running Poolside Laguna S-2.1 UD-Q5_K_S same 3090 + 64GB RAM same hybrid offload q8_0 KV 166k ctx Q4 was ~73 GB. Q5 is ~83 GB. More bits on disk. Hopefully this one is better, We’ll see if it’s worth it.
AJ13,341 görüntüleme • 3 ay önce

Every big MoE model I test on my 3090 is basically to see how good it is compared to Qwen3.6-27B. "How good is it compared to Qwen3.6-27B?" is the question I get asked the most, and that shows how much people love and, more importantly, trust the 27B 27B is battle tested. It has proven its worth and has nothing left to prove. Now it just needs to make way for the next 27B. Here’s an example video: a Three.js task, same prompt, not oneshot but two shots for all models. All ran on my 3090 + 64GB VRAM Qwen-3.6-27b fully on vram while others have experts offloaded to system ram. Pinned on the left: Qwen3.6-27B-Q5 On the right, i'm switching between Inkling-Small-Q4_K_M DSV4-Flash-0731-UD-IQ1 Ling-3.0-Flash-Q4_K_M Qwen3.6-27B nailed it. The effects and visuals are just spot on.The other models got some things right, but not everything. Second place would have to go to Inkling-Small while DSV4-Flash-0731 and Ling-3.0-Flash failed to capture the full requirements. This is also why we’re all excited for Qwen Qwen3.8-27B!
AJ11,670 görüntüleme • 3 ay önce
Daha fazla içerik yok.