antirez
@antirez · Sicily, Italy · joined 06 May 2007
Reproducible bugs are candies. I like programming too much for not liking automatic programming.
80 057Followers
799Following
43 633Posts total
922.4KViews on collected posts
Ultimi post
Cerchiamo nuovi post…
I ran an extensive benchmark against DeepSeek v4 Flash and GLM 5.3 Flash Q2, Q4 and mixed quants. Those are the results obtained. Mix of (hard-ish) benchmarks on cybersecurity, math, QA, ... https://t.co/43Kjoh2TmD
6.9K views · 89 likes
· 5 reposts · 10 replies
Open on X →
Do you believe me now, after the Astra issue, that Artificial Analysis numbers are broken?
16.4K views · 210 likes
· 6 reposts · 20 replies
Open on X →
We don't give a !(@$# about what the new model can generate with three.js. We want to know if you had a blocking problem X in software Y that many rounds of the old models didn't fix, and the new model just released improved the situation.
14.2K views · 520 likes
· 35 reposts · 31 replies
Open on X →
P.S. wrong URL! The steering file is at https://t.co/vuvQAk5joH
10.3K views · 44 likes
· 0 reposts · 3 replies
Open on X →
Asymmetric access to LLMs for cyber security workis bad. I decided to release the steering vector that you can use with DwarfStar --dir-steering-file <file> and --dist-steering-ffn (try 2, 3, 4 based on refusal) to make DS4F comply. https://t.co/5mJYHkX0Es
27K views · 226 likes
· 16 reposts · 10 replies
Open on X →
@antirez V4 Flash Vision uncensored is better https://t.co/UeHfCY4y53
1.9K views · 24 likes
· 0 reposts · 2 replies
Open on X →
@antirez It is such a good model for local usage.
2.9K views · 23 likes
· 0 reposts · 0 replies
Open on X →
@antirez I'm also a fan of Deepseek-V4-Flash. It is one of those rare open source models, which can actually do well in less popular languages (not English or Chinese) + 1 million context window and still relatively smart at 500k-600k is amazing. I do use it via official DeepSeek
1.7K views · 18 likes
· 1 reposts · 0 replies
Open on X →
P.S. for me DeepSeek v4 Flash is still the king of local inference, regardless of other great models released that could beat it in specific benchmarks and when many (many!) tokens are used. Especially now that DS4F has vision. What a model...
71.3K views · 878 likes
· 33 reposts · 87 replies
Open on X →
I'm optimizing tensor parallel execution across two MacBook M5 Max systems using RDMA. This is how fast DeepSeek v4 Flash Exp Vision can run in MXFP4 (full precision) right now *without* DSpark, so this is the min speed you get in practice. https://t.co/pTL0lTOxDy
18.2K views · 289 likes
· 22 reposts · 21 replies
Open on X →
If you develop Metal kernels using remote machines, beware that the freaking screen saver burns GPU cycles and must be replaced with solid or disabled to have reliable benchmarks. My kindly donated two m5max systems are in a air conditioned room and I was bite by this issue.
13.8K views · 188 likes
· 4 reposts · 6 replies
Open on X →
DeepSeek V4 Flash Vision. Abliterated. GGUF. Run it yourself. For free. 🐳
We just released DeepSeek V4 Flash Vision Abliterated GGUF.
→ Abliterated weights
→ Vision + text
→ GGUF for local inference
→ Run it on your own machine → Free & open weights
Want the standard
239.6K views · 505 likes
· 24 reposts · 23 replies
Open on X →
You see? Those are not complicated concepts. However sometimes papers, and almost always LLMs at this point, are not able to explain stuff in simpler terms, even when the underlying ideas are simple. Nobody apparently cares much, but I hope the dislike of Opus 5 jargon will teach
22.3K views · 132 likes
· 6 reposts · 5 replies
Open on X →
Now learned n-grams do this at the next level, like if single tokens were single characters in our first example. And they do it for many common two / three tokens groups. Since bi and especially trigrams are too many, a compressing hashing is used for indexing.
5.4K views · 43 likes
· 0 reposts · 2 replies
Open on X →
One way to understand why this is useful: think to some LLM that gets input sequences as single chars. The first layers will spend a lot of space + compute in order to group things like a,p,p,l,e to their meaning of apple. Learned embeddings solve this.
3.7K views · 38 likes
· 0 reposts · 1 replies
Open on X →
This way the token "York" of " bla bla bla New York" will be enriched with the "New+York" meaning embedding, allowing the next layers to process more directly the higher level concept.
3.8K views · 31 likes
· 0 reposts · 1 replies
Open on X →
So what happens is that at layer 1 you take the last two and three tokens, get 16 small vectors from the n-gram table, you compose them into a larger vector that is gated in order to enrich the current token at layer 2.
4K views · 34 likes
· 0 reposts · 3 replies
Open on X →
The idea is that this way you save a lot of weights on the model backbone, that need to remember how to process associations like "New York", "Computer Science", "Three Body Problem", and so forth. The ~2500 vector retrieved from 51B database is sparse to deal with collisions.
5.8K views · 48 likes
· 0 reposts · 2 replies
Open on X →
In case you don't know, the Qwen 3.8 Flash Next n-grams "enrich" the current token representation at layer2 using a 51B table and a gate to understand how much we want to pass. Layer 1 processing gives you the time to get the 16 small vectors from the SSD, masking loading.
8.7K views · 83 likes
· 2 reposts · 2 replies
Open on X →
Given that 51B of n-grams can stay on the SSD disk, Qwen 3.8 Flash Next with a 2 bit quantization could be the best DwarfStar bet for a 64GB MacBook local inference top experience. I understand @ivanfioravanti can spend some time with the implementation. Let's see what happens.
44.1K views · 542 likes
· 14 reposts · 30 replies
Open on X →
How do you trust a metric where 1st and 2nd place are those? GPT 5.6 Sol is in the real world a *fundamentally* strong model than Opus 5. It means, of course, that all the rest of the chart is partially shuffled. Take it as one of the many signals. https://t.co/EpYY9CgOtT
55.2K views · 460 likes
· 22 reposts · 50 replies
Open on X →
@antirez There is a little mistake in the ePub file.
Look at the header. https://t.co/xjXUfnLl5H
1.3K views · 1 likes
· 0 reposts · 1 replies
Open on X →
Please note that this will be also a proper release in Italian language. I'm just refining the last details.
4.1K views · 6 likes
· 0 reposts · 0 replies
Open on X →
My second short story release in English is ready: Tales of Illustrious Computer Scientists: Iola Varga, nun and computer scientist. https://t.co/i1Xp0REuWh
321.9K views · 230 likes
· 13 reposts · 17 replies
Open on X →
Rispetto ad account della stessa dimensione
23 post degli ultimi 90 giorni, accanto alla fascia di 10K–100K follower. arriva a molti, ma pochi di loro reagiscono.
Visualizzazioni mediane10 263questo account924mediana per 10K–100K
Copertura, %12.82%questo account3.62%mediana per 10K–100K
Interazione, %1.06%questo account1.52%mediana per 10K–100K
| Metrica | Questo account | Mediana per 10K–100K | Rapporto |
|---|---|---|---|
| Visualizzazioni mediane per post | 10 263 | 924 | 11.1× |
| Copertura (visualizzazioni ÷ follower) | 12.82% | 3.62% | 3.54× |
| Tasso di interazione | 1.06% | 1.52% | 0.70× |
Altri account di questa fascia → Confronta con un altro account → Come sono costruiti questi parametri →
Growth & engagement
How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.
Views per post
1.7K30 Aug
239.6K2 Sep
16.3K
13.8K
18.2K
71.3K
1.7K
2.9K
1.9K
27K3 Sep
10.3K
14.2K
16.4K
6.9K
Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.
Engagement rate per post
1.38%30 Aug
0.23%2 Sep
1.89%
1.44%
1.84%
1.41%
1.10%
0.80%
1.40%
0.94%3 Sep
0.46%
4.19%
1.44%
1.53%
Reactions — likes, reposts, replies and quotes — divided by views. Median for 10K–100K accounts is 1.52%.
What the audience does
Likes80.9%4 987 in total
Reposts3.3%203 in total
Replies5.4%335 in total
Quotes0.6%38 in total
Bookmarks9.7%598 in total
Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.
The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.
Account simili
hodlonaut ✓
@hodlonaut
160.1K followers
· 59.3K posts
船瀬俊介の船瀬塾
@funasejuku
160.1K followers
· 4.3K posts
الاهلي السعودي
@AlahliiNow
160.1K followers
· 14.8K posts
yarım kalanlar
@arlanmaz
160.1K followers
· 5.4K posts
YourFavoriteGuy ✓
@guychristensen_
160.1K followers
· 16.2K posts
İlker Ayrık / Pervasız
@pervasiz
160.1K followers
· 1.4K posts
生涯収支マイナス５億円君
@maiokux
160.1K followers
· 599 posts
Noirchick In Old Hollywood ✓
@Noirchick1
160.1K followers
· 40.4K posts