Nid All | AI art and more 🤖

@bindureddy The real question is, will it be better for agentic workflows?

Gemini is over due. Expect a really good Flash model Something that beats GPT 5.5/Opus 4.7 medium but is considerably cheaper and faster

Starting today, TinyFish Web Search and Fetch are free. For every dev and agent. Across the galaxy. No credit card. Generous rate limits. Grab your API key from tinyurl.com/bdzer5x6 May the web be with you.

@jatinkrmalik @opencode DS V4 pro is a tier 3 model

Bookmark this if you're on @opencode Go!👇 I mapped out exactly how to use all 14 SOTA models without hitting your rate limits, which agent gets which model, how fallback chains work, and when to escalate between tiers. Steal my setup w/ the full guide: @jatinkrmalik/opencode-go-oh-my-openagent-the-complete-guide-to-sota-model-routing-without-hitting-limits-49fdc8cb3417" target="_blank" rel="nofollow noopener">medium.com/@jatinkrmalik/…

@teortaxesTex It’s V4 preview after all.

Many are saying V4 (speaking purely of Pro) is strange. It is often unreliable (hallucinates, falls apart) and often rock solid. At its best, it's way more "Frontier" than other 1T class Chinese open models. And yes, *good* at debugging. It's like something not fully merged.

@Hunter171270 Combining world models, image and video models, and MLLMs is like collecting the Infinity Stones, and Google would do that sooner or later.

Veo3 and Sora2 were GPT-3.5 moment. Seedance 2.0 was GPT-4 moment. I really hope we're close to o1 and o3. First window is Google I/O.

@teortaxesTex We know who will do that soon.

Bro thinks people finetune Kimi K2.6 alas… I wish bro, I wish…

@teortaxesTex I have consumed 150M tokens and had no major issues. The throughput is consistent compared with @Zai_org GLM 5.1, which is painfully slow.

Hero. but I've just noticed, V4 was an end of an era. For very long, DeepSeek was proud of NOT having explicit rate limits. They literally had a «rate limit» section in API docs just to brag about it. They really are inference-compute-limited now. Largely by such heroes, I guess.

GitHub if it were made in Algeria 🇩🇿

@ivanfioravanti @angeloskath The KV cache memory usage is insane whale bros did an amazing job.

Another preliminary MLX DeepSeek 4 Flash context benchmark (up to 32k context), to check the status of the PR on mlx-lm now that @angeloskath is on it 💪 Things are progressing well as you can see below. 4bit is pretty fast and sustained across contexts. Hardware: Apple M3 Ultra 512 GB 0.5k pp 404 tg 32 t/s mem 155.2GB kv 0.02GB 1k pp 460 tg 32 t/s mem 155.5GB kv 0.02GB 4k pp 437 tg 30 t/s mem 157.7GB kv 0.04GB 8k pp 425 tg 29 t/s mem 157.7GB kv 0.07GB 16k pp 350 tg 29 t/s mem 158.8GB kv 0.12GB 32k pp 316 tg 28 t/s mem 161.3GB kv 0.23GB Batch TPS: b1 30 b2 43 b4 60 b8 80 b16 116 b32 155

@intheworldofai @arena gemini ?

🚨 NEW Stealth Model "kartoffeln" in @arena - Pretty good model with great webdev capabilities Kartoffeln is the German plural noun for "potatoes". OpenAI's GPT-5.5's internal codename was "spud", a common term for a potato👀 kartoffeln could be another OpenAI Model!

Keep getting rate-limited by Claude, so I tried out DeepSeek V4 for the first time. After 10M+ tokens, holy crap the cost is ... 🤯

@TeksEdge 3.1 flash

Gemini 3.1 Flash Lite incoming!

@pseudokid DeepSeek V4 Pro : Backend

Kimi K2.6 - frontend Deepseek V4 Flash - cheapest all-rounder, cache king, 1M Mimo V2.5 - excellent vision, 1M Qwen 3.6 Plus, Mimo V2.5 Pro - decent all-rounders, 1M MimiMax M2.7 - decent backup All current gens are way better at using the terminal than their prev gen

Trying to find a solution for the long verbose CoT of Qwen 3.5 2B model using the verbosity of DeepSeek V4 Pro Max @teortaxesTex

@gnukeith V4 Pro by far

Best Chinese model right now? I'm using Kimi K2.6 myself

@Rydo111 @0xSero @opencode I use it with openclaw too it’s good for general purpose stuff

@n1d_all @0xSero @opencode I'm using opencode as well, with oh-my-openagent. There are multiple agents that you can set different models to

I LOVE Deepseek-v4-flash, incredibly reliable and capable, logical. It's lacking in frontend but I have MiMo for that. I would recommend any company spending 100k+ a year on AI to purchase 8-10~ 6000s and have a few of the works to have them blind test these models for work.

@Rydo111 @0xSero @opencode

@n1d_all @0xSero Which agent do you set DS on?

@Elaina43114880 It gonna take 6-8 hours, cool do it now xD