Ollie Codes

Tech geek 🤓 | NextJS & SEO enthusiast | Dog lover 🐾

[FHILY👑 @Oluwaphilemon1]

Qwen3.8-27B-NVFP4 just passed 1.3M downloads. And the numbers explain why.

This isn’t just another 4-bit quant.

27B dense | Vision | 256K context | NVFP4 | MTP | Apache 2.0

The Blackwell results are wild:

→ RTX 5090: ~110 tok/s with vLLM + MTP

→ DGX Spark: 132.1 tok/s NVFP4 vs 99.5 tok/s FP8

→ 2× RTX 5070 Ti: ~54–64 tok/s

And the model drops from 30.9GB to 23.4GB.

The catch: NVFP4 is built for NVIDIA Blackwell. For Mac, AMD, or Intel, GGUF is still the better route.

[Marc Lou @marclou]

- https://t.co/8rZYnV6Vf6 hits $4K/day
- https://t.co/tTyP0BVJh7 all-time @ $44K/month
- Emergency in Japan 🇯🇵
- MARKETING OR DIE
- https://t.co/sByPTnJwEw is stale

I hope you enjoy the VLOG!

[Paul Mit @pmitu]

Indie hacking will die once people start generating their own apps with a single prompt

[Jukan @jukan05]

Meritz Securities: According to our channel checks, Japan’s Ajinomoto has recently informed customers that, due to an acute supply shortage, it will prioritize allocations for AI and high-end applications. As of 2Q26, its ABF film production lines are already operating at full capacity, while meaningful contributions from new capacity additions will take considerable time to materialize, leaving limited scope to expand supply in the near term. Consequently, limited ABF film volumes are expected to be allocated primarily to key AI customers and high-end products, further concentrating supply among leading substrate manufacturers with proven mass-production track records and stable yields in high-end ABF substrates. This could strengthen the leading players’ bargaining power within the supply chain and provide greater scope for price increases. Ajinomoto’s allocation strategy prioritizing AI and high-end demand is therefore expected to further widen the gap between leading and lagging players.

[Wes Roth @WesRoth]

Actual DeepSWE run for Ox Alpha: ~63%.

Not the crazy 80% from the earlier subset test.

But apparently it’s still VERY good:

→ strong coding
→ handles subagents well
→ good on long complex tasks
→ roughly Sol Medium-tier

Main weaknesses: can leave dead code behind and gets pretty slow at high reasoning.

[Mi Rony @mironydev]

Another one:

“Too expensive” can be just as useful as “I want this.”

Find people who want the solution,

but aren’t happy with the existing options.

[Tommy D. Rossi @__morse]

@damonchen finally

[Aysha @Aysha__Mohamed]

The UAE’s partnership with Somaliland is visible on the ground.

From Berbera Port to the Berbera–Hargeisa corridor, UAE investment has helped build infrastructure that connects cities, reduces transport costs, supports businesses and opens new trade routes toward Ethiopia and East Africa.

[Vijar Kohli @VijarKohli]

Me: Please repackage this concept.

Gemini: It's a good idea, but not a great idea.

Me: 🤦‍♂️ well then, make me one billion dollars

[Adit_Yah ☄️ @Adidotdev]

🚨 Stop calling Ox Alpha Gemini already

Full DeepSWE just dropped: 63% on all 113 tasks.

That’s above Opus 4.8.
Right behind GLM-5.3 and GPT-5.6 Sol.

The early 80% was noise from a tiny 10-task subset. This is the real score.

And the identity? Pretty much done.

Someone forced the server to name itself. It threw a Java stack trace with Zhipu’s own internal class. Tokenizer matches GLM-5.3 on basically every probe. Video encoder behaves exactly like GLM-5V-Turbo. Public GLM-5.3 is text-only - this one sees video.

So no… it’s not Gemini.
It’s almost certainly Zhipu’s unreleased multimodal Flash version of GLM-5.3.

A free Chinese Flash model sitting this close to Sol and Opus 4.8 is kind of insane.

What the hell is happening over there.

[Denis / Backlinks - Digital PRs 🪐 Passive Sphere @PassiveSphere]

12 days till TOP 5 on Google

While everyone is obsessed with being high on the leaderboard

I keep publishing blog posts on my SaaS site

- you get traffic that doesn't convert
- I get highly targeted traffic that converts

Build backlinks and your blog posts will rank well too!

[Haider. @haider1]

curious why METR hasn't updated its model data since May

though i think this metric might become less useful anyway cuz models are just too jagged across tasks

the next step should be continual learning -- models you can actually teach and improve through human feedback

[VraserX e/acc @VraserX]

The next model war may not be about intelligence at all.

It could be about how long a model can operate autonomously before a human has to rescue it.

That metric would matter far more to businesses than another benchmark leaderboard.

[Tommy D. Rossi @__morse]

the code of this app was ported from @localhost_5173 's Waku. the code is in the repo

it has support for instant HMR which lets you develop much faster compared to just Rust and GPUI

styling is very similar to HTML + React so agents are good at it. this app was just one shotted from a screenshot

https://t.co/swyoL6zCuR

[Haider. @haider1]

wondering when Astra / Fable 2 actually drops

oai and anthropic have hit the brakes so hard i'm starting to wonder if either will even release their best models publicly anymore, or if we'll have to depend on china for that

1-2 months is fine

but 3+ months between releases would suck

[AFT DAVID @aft_david]

Turns out the algorithm wasn’t the problem.

We just spent years optimizing the wrong inputs.

[Doğukan @DogukanUrker]

Qwen3.8-27B on a single RTX 3060 12GB. Best 2-bit setup I've found for a 12GB card. (config below)

Same GGUF, one flag apart:
MTP on -> 40 tok/s, 110K ctx
MTP off -> 23 tok/s, 165K ctx
Prefill barely moves (470 vs 493). It's decode speed vs context, nothing else.

Weight is @atomic_chat_hq AD-IQ2_XXS (8.98 GB), not @UnslothAI's smaller 7.27 GB -- the MTP head is inside the GGUF instead of a sidecar. Unsloth's chart plots size "with removal of MTP"; add the sidecar back and the gap mostly closes.

Draft acceptance 0.57-0.71, mean len 2.7-3.1.

I run reasoning_effort low. medium is fine for some tasks, but the default xhigh doom-loops at 2-bit. (effort level benchmark in the replies)

-m Qwen3.8-27B-AD-IQ2_XXS.gguf -ngl 99 -c 110000 -fa on --jinja -np 1 --spec-type draft-mtp --cache-type-k q4_0 --cache-type-v q4_0 -b 1024 -ub 512 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --reasoning on --chat-template-kwargs '{"reasoning_effort":"low"}'

Drop --spec-type draft-mtp and set -c 165000 for the long-context variant.

[Simplifying AI @simplifyinAI]

turn on prompt caching, save money and time. that's the pitch every provider makes.

it's only half true. a new paper tested OpenAI, Anthropic, and Google's flagship models across 500+ agent sessions, and the cost side holds up every time, 41 to 80% cheaper no matter how you cache. the speed side doesn't.

cache everything by default and the provider keeps writing fresh entries for tool calls and results nobody reuses. one model actually came out slower than not caching at all.

"naively enabling full-context caching can paradoxically increase latency"

the fix is boring but it works: cache the system prompt, cut the boundary before the dynamic stuff starts. same speed gains, none of the downside.

check the article below for the full breakdown on caching, context, and everything else quietly running up your bill 👇

[Haider. @haider1]

woah, apparently this stealth model "Ox-Alpha" just mogged Fable 5 and GPT-5.6 Sol on DeepSwe

i haven't tried it on large-scale architecture work yet, but as an orchestrator, executor, and reviewer, it's been more than capable

definitely a step above 5.6 luna

my guess is MiMo

[Chubby♨️ @kimmonismus]

wtf is happening today: DeepSeek has launched DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model built for agents that need to see.

And its performance on visual-agent benchmarks moves close to or even outperforms Opus 4.8.

Again: this is the Flash model, the small one!