Currently building in the shadows, all day long - Former chair of @Wikimedia & @Wikimedia_Fr - he/him

Paris
Joined March 2007
Christophe Henner retweeted
Is JD Vance seriously suggesting that immigrants have brought a culture of revolution and protest TO FRANCE? @jonstewart
187
1,685
100
8,222
235,184
Christophe Henner retweeted
zml/llmd is now faster than vLLM on DeepSeek-V4.1-Flash on 4xGB300. The remaining 30 tok/s are already accounted for and will be gone very soon. And most of those optimisations generalize to other models. Little by little.
6
4
41
2,298
Christophe Henner retweeted
Ummm guys, I took @googlegemma Embedding 2 model and train it on 3,000hrs of Vietnamese audio. IT REACHED SOTA. All it took is 2 hours on 8xH100
62
81
11
1,287
68,889
Christophe Henner retweeted
Like OpenAI models, Haiku is cheaper on short context, Add under modelSettings in ~/.claude/settings.json json { "modelSettings": { "claude-haiku-5-5": { "autoCompactWindow": 100000 } } } You will save a shitload of quota without doing anything.
Replying to @ClaudeDevs
For prompts under 100K tokens, it's $0.10 / $0.50 per million tokens with cache reads at $0.01. Above 100K tokens, $0.50 / $2.50 per million, with cache reads at $0.05.
1
2
1
190
Christophe Henner retweeted
FYI
ZML already runs on Tenstorrent, Qualcomm, Intel and FuriosaAI btw
6
56
1,911
Christophe Henner retweeted
For context, the May 68 riots were the most violent ones. Nothing came close since then. In France, they are considered a fight for freedom in common culture.
5
1
11
3,440
Christophe Henner retweeted
Some personal news: I’m starting at @WIRED as their new senior politics writer!
47
19
2
602
19,195
Christophe Henner retweeted
Our first ZML/LLMD release with a frontier model ought to be DeepSeek. And we did. Enjoy!
1
29
3,032
Christophe Henner retweeted
1
1
1
20
3,303
Webinaire gratuit : Copywriting. Nous serons avec Mehmet Gul, @selim, Axel Thoreux et moi-même ce vendredi à 15h15 connectés pour débattre de copywriting aujourd’hui. La session sera animée par @Elliott_bobiet. Lien pour y assister : seo-summit.com/webinaire/?id…
2
3
6
378
Christophe Henner retweeted
18300 tok/s at bs=1024
update: Qwen 3.8 27B BF16 now at 17500 tok/s on 4xGB300
1
1
2
26
3,948
Christophe Henner retweeted
ArseneLupin il décide vite, il décide bien Bon vous avez tous vu la hype JEV, vous me connaissez j'ai direct voulu faire un modèle équivalent, il s'appelle ArseneLupin, grosso modo 48h de training plus tard voici ce que donne la V1 (la V2 à venir dans la semaine)
10
12
2
105
9,406
Christophe Henner retweeted
taking that opportunity to highlight the following: 1/ Cuda Tile (TileIR) doesn't get you 100% speed, sadly 2/ Cute-DSL (CuTeIR) does get you 100% speed 3/ ZML supports both inline inside the graph, JIT :) 4/ or Triton, or Gluon and all of this with zero python
Introducing CUDA Rust! CUDA Rust lets you write GPU kernels natively in Rust, not just launch them from it. Two paths: cuda-oxide for SIMT kernels compiled to PTX, and cutile-rs for Tile-based programming on stable Rust. Both can catch aliasing errors at compile time. Technical blog: nvda.ws/4hm1bHS
3
6
1
114
10,288
Christophe Henner retweeted
we now have deepseek v4 and GLM running faster than both vllm and sglang lots of features work to do, but here we are
4
5
3
68
6,165
Christophe Henner retweeted
Aujourd'hui est une journée importante pour @Ibou_io avec la sortie d'une nouvelle verticale : ibou actualités. Je félicite toute l'équipe qui s'est bien arrachée pour vous fournir un outil très pratique pour vous pointer les bonnes actus à tout moment. actualites.ibou.io/
3
12
1
34
2,204
Christophe Henner retweeted
Après 13 années mémorables chez PlayStation je suis très heureux de vous annoncer que je rejoins SHIFT UP en tant que Head of Global Communications ! Hâte de piloter les campagnes mondiales pour les pépites de M. Kim Hyung Tae, M. Shinji Mikami et leurs équipes !
1,531
131
407
2,076
1,679,001
Christophe Henner retweeted
En plein mois d'août on s'est dit qu'on allait faire un podcast entre nous. Avec les congés c'est plus simple, et ça permet de répondre à plein de questions qu'on nous a posé sur @Ibou_io youtube.com/watch?v=LbczK-zo…
2
1
8
1,403