Micke CORNIER

[Micke CORNIER @micke_cc]

@ELISIUMcity @Alibaba_Qwen A 3.8 - 80b a3b would be wonderful

[Micke CORNIER @micke_cc]

@MistralDevs @mistralvibe @MistralAI
openAi a publié gpt-oss 20 et 120b, des MoE open sources. Google a publié gemma 4, avec notamment 26ba3b.
Alibaba l'excellent qwen 3.6 35ba3b.

Vous qui aviez été parmi les premiers à faire un MoE avec Mixtral, pourquoi ne proposez vous pas un Mistral MoE équivalent ?

[Micke CORNIER @micke_cc]

@Photomaton_Off prix doublé en 6 ans pour les photos officielles :
5€ (2020) → 8€ (avr. 2026) → 10€ (juin 2026)
+25% en deux mois. Et le résultat n'est même plus une photo papier réutilisable, juste un code à usage unique. Une explication ?

[Micke CORNIER @micke_cc]

@googlegemma Thank google,
Waiting a diffusion llm since 1y ! (Since Mercury llm announcement)

[Micke CORNIER @micke_cc]

@lmstudio This test shows :
- Windows Taskmanager bug with unified memory, 20go can be loaded for GPU and NPU, there is no limitations.
- NPU doesn't support few specifics tensors pipelines, like DeltaNet.
- OpenCl doesnt support mxfp4.

[Micke CORNIER @micke_cc]

@lmstudio
https://t.co/rgDRom4IS1
When will OpenCL and Hexagon runtime support come to LM Studio?
I tested it on my Surface Pro 11 with Snapdragon—GPT-OSS-20b runs well on Hexagon NPU (8.6 tok/s), but Qwen 3.5 with Deltanet attention aren't yet supported on Hexagon/OpenCL...

[Micke CORNIER @micke_cc]

A Vibe-coded UI using OpenCL and Hexagon with Llama.cpp The NPU shows solid performance with GPT-OSS 20B, but doesn't support Qwen 3.5 and Gemma 4 yet.
8 tps is lower than 12 tps on CPU, but power consumption is vastly different and prevents system freezes during inference

[Micke CORNIER @micke_cc]

Hi @lmstudio — any ETA on OpenCL runtime support for llama.cpp in LM Studio?

I built and tested the llama.cpp OpenCL version locally: my Snapdragon X Elite uses it correctly, though performance is currently similar to CPU on oss-20b.

[Micke CORNIER @micke_cc]

@Microsoft @LMStudioAI @lmstudio

[Micke CORNIER @micke_cc]

@Microsoft @LMStudioAI
Nov 2024 blog said LM Studio was “powered by ONNX Runtime” and could use Snapdragon NPUs on Copilot+ PCs.

Dec 2025: public LM Studio is still GGUF/MLX only (no ORT GenAI/NPU on X Elite). Demo-only? Any roadmap?

[Micke CORNIER @micke_cc]

@Microsoft @LMStudioAI Was that demo/prototype-only? Any public roadmap or requirements for ONNX support (EP: QNN/DirectML, model formats, conversion pipeline)?

[Micke CORNIER @micke_cc]

@Microsoft @LMStudioAI It’s Dec 2025 and the public LM Studio still focuses on GGUF (llama.cpp) + MLX, with no visible ONNX/ORT GenAI pipeline or NPU option on Snapdragon X Elite.

[Micke CORNIER @micke_cc]

@Microsoft @LMStudioAI In a Nov 2024 Microsoft blog, LM Studio was described as “powered by ONNX Runtime” and able to leverage Snapdragon NPUs on Copilot+ PCs.

[Micke CORNIER @micke_cc]

@Microsoft @LMStudioAI https://t.co/YW9ibIwBpP

[Micke CORNIER @micke_cc]

I just cast my VOTE for CLAIR OBSCUR: EXPEDITION 33. Who will you vote for? #TheGameAwards @TheGameAwards https://t.co/b3XQwsqETN

[Micke CORNIER @micke_cc]

@Microsoft @MicrosoftEdge j’en ai marre. J’utilise Edge depuis des années (boulot inclus), mais c’est la fois de trop : j’ai remis Google en page « nouvel onglet » il y a 2 semaines et, après une MAJ Windows, votre page revient. Arrêtez de changer nos réglages.

[Micke CORNIER @micke_cc]

@OpenAI After the release of the amazing gpt-oss, I was expecting a technical Game changer.

Token Diffusion have a great potential and are faster. (Test Mercury or the gemini Diffusion).

[Micke CORNIER @micke_cc]

@OpenAI Still a AR... not a token Diffusion model :/

[Micke CORNIER @micke_cc]

@GroupamaStadium @Imaginedragons @lubodomo Première partie horrible, trop de réverbération !

[Micke CORNIER @micke_cc]

@GroupamaStadium première partie inaudible, une honte...
Trop de réverbération !