NIKAN WYSTAF

AI nerd.
Graphic Designer and a little bit Programmer.

[NIKAN WYSTAF @WYSTAF]

I'm really excited for the upcoming Qwen4 models (still my favorite LLM family).

My Qwen4 projections:
27B xhigh ≈ 45 (GLM-5.3 Max)
Plus ≈ 46-47 (MiMo-V2.6-Pro)
Flash ≈ 48 (Muse Spark 1.3)
Max ≈ 51 (GPT-6.1 Sol)
Max Update ≈ 56 (Sonnet 5.5)

Explanation in comments 🔻

[NIKAN WYSTAF @WYSTAF]

@Alibaba_Qwen @ArtificialAnlys @QwenDevs If the same pattern holds on top of the initial Max score, Qwen4 Max (Updated) could reach around 56, putting it right next to the current Claude Sonnet 5.5.
Curious to see how close the real numbers will land when Qwen4 drops.

But it's still a guess .

[NIKAN WYSTAF @WYSTAF]

I'm really excited for the upcoming Qwen4 models (still my favorite LLM family).

My Qwen4 projections:
27B xhigh ≈ 45 (GLM-5.3 Max)
Plus ≈ 46-47 (MiMo-V2.6-Pro)
Flash ≈ 48 (Muse Spark 1.3)
Max ≈ 51 (GPT-6.1 Sol)
Max Update ≈ 56 (Sonnet 5.5)

Explanation in comments 🔻

[NIKAN WYSTAF @WYSTAF]

@Alibaba_Qwen @ArtificialAnlys @QwenDevs And the most interesting part:
Just like Qwen3.8 Max got a strong post-release update (0902) that pushed it from 40 to 45, I expect Qwen4 Max to also receive a meaningful update 1-2 months after launch.

[NIKAN WYSTAF @WYSTAF]

@Alibaba_Qwen @ArtificialAnlys @QwenDevs Based on this consistent +11 jump, my projections for Qwen4:
• Qwen4 27B xhigh ≈ 45 → around GLM-5.3 Max
• Qwen4 Plus ≈ 46-47 → MiMo-V2.6-Pro / Grok 4.7 level
• Qwen4 Flash ≈ 48 → Muse Spark 1.3 Max
• Qwen4 Max (initial) ≈ 51 → GPT-6.1 Sol level

[NIKAN WYSTAF @WYSTAF]

@Alibaba_Qwen @ArtificialAnlys @QwenDevs Found a clear +11 pattern on the Artificial Analysis Intelligence Index across :
• Qwen3.5 27B (23) → Qwen3.8 27B xhigh (34)
• Qwen3.7 Plus (25) → Qwen3.8 Plus Estimated (38)
• Qwen3.7 Max (29) → Qwen3.8 Max (40)
• Qwen3.8 27B xhigh (34) → Qwen3.8 Max 0902 (45)

[NIKAN WYSTAF @WYSTAF]

@ItsmeAjayKV Totally Agree, Thanks for mentioning.

[AJ @ItsmeAjayKV]

Qwen has been the biggest and strongest enabler of Local AI for consumers.

As i speak to more and more local AI users with less than 12GB VRAM, the situation feels pretty sad.

I'm specifically talking people with around 4 - 12 GB of VRAM or Unified memory, 3080, 3060, 4070, 4080 and similar setups.

They're unfortunately being left out.

No. Asking them to upgrade is not fair, esp in this economy?

As someone who started local AI with only a 12GB 3060, i understand them, Qwen3.6-35B-A3B is the only reason i'm here today.

From what i've seen, many users are still sticking with Qwen3.6-35b or gemma4-26b, or some other older MoEs because they simply don't have many newer options they can run well.

@Alibaba_Qwen, for future please releases, please consider a
- 9B model
- 35B model,
- or some new substiture MoE that can give more value to low vram users.

Maybe this can start from Qwen4.0 🙏

[NIKAN WYSTAF @WYSTAF]

@TokenHarborAI Thanks Tokenharbor, you helped me a lot 🫡

[NIKAN WYSTAF @WYSTAF]

I’m going to start releasing the projects one by one right here over the next few weeks.
If it looks like a bunch of things suddenly appeared in a short time, now you know why.
Also… I still have a relatively light game I’ve been working on for two years :)

[NIKAN WYSTAF @WYSTAF]

I have a couple of bad habits I’m actively trying to fix.

First: whenever multiple ideas hit me, I start working on all of them at the same time.
Second: I’m a perfectionist. I refuse to ship anything until it feels “good enough” to me.
Those two things...

More in the replies🔻

[NIKAN WYSTAF @WYSTAF]

Perfectionism just wouldn’t let me hit that push button.
I started coding more or less in 2023, but my GitHub only shows activity from 2026, because I almost never pushed anything.

[NIKAN WYSTAF @WYSTAF]

Those two things combined mean most of my projects stay private for way too long.

Classic example: for years I barely used git.
I’d keep coding until I had a “perfect” version, then push.
Result? Full-time work, overtime, almost no sleep… and almost no commits on GitHub.

[NIKAN WYSTAF @WYSTAF]

Today marks exactly one month of full-time work on this.
Most of the time went into Kaggle TPU limits (grateful for the free access) and my almost zero experience.
Treated it as a learning project and picked up a lot about JAX, KV caching, sharding and more. It’s almost done now.

[NIKAN WYSTAF @WYSTAF]

I have a couple of bad habits I’m actively trying to fix.

First: whenever multiple ideas hit me, I start working on all of them at the same time.
Second: I’m a perfectionist. I refuse to ship anything until it feels “good enough” to me.
Those two things...

More in the replies🔻

[NIKAN WYSTAF @WYSTAF]

@GaelBreton Gemini 3.8 Flash Extended :

[NIKAN WYSTAF @WYSTAF]

@AmirStillAlive سرت سلامت ❤️

[NIKAN WYSTAF @WYSTAF]

@imansh237 @AmirStillAlive منم گول خوردم داداش 😁

[NIKAN WYSTAF @WYSTAF]

First time seeing ads in the ChatGPT app and I'm not against it.
They gave us GPT-5.6 Luna for free, and that definitely isn’t free for them to run.
Ad revenue can help fund further development and keep improving the models.

@OpenAI
@ChatGPT

[NIKAN WYSTAF @WYSTAF]

@Ansh14225311 🫡😄