Just discovered jobbench evaluation, I like it.
job-bench.github.io/#Leaderboard
=> Opus 5.5 dominates.
By the way, @nanbeige 4.3 33B may be worth trying...
huggingface.co/Nanbeige/Nanbe…
job-bench.github.io/#Leaderboard
=> Opus 5.5 dominates.
By the way, @nanbeige 4.3 33B may be worth trying...
huggingface.co/Nanbeige/Nanbe…
0
1
2
@KuptoKosmos En même temps si tu coupes les réseaux y'a pu d'faille 🤓
Il est fortement conseillé d'utiliser l'Internet assis, et pas debout.
Il est fortement conseillé d'utiliser l'Internet assis, et pas debout.
0
0
1
Dans la voiture tu te brises les cervicales, c'est la plus belle mort que je puisse imaginer.
😁😁😁
😁😁😁
0
0
0
@Aido_Tweet @AlmanaxWinch Dans le pipe 3D ça donne ça
Y a une grosse perte de perf mais avec le framegen et les futures cartes ça passera creme
Y a une grosse perte de perf mais avec le framegen et les futures cartes ça passera creme
1
0
0
I'm totally blown away by Strata from @coldniko
Installation process is so smart and easy.
Performances are incredible (IQ3S - KV Q8 here)
Monitoring interface is great...
BRAVO!
2x RTX 3090
128 GB DRAM (8 channels)
Threadripper 5965WX 24 cores
PCIe4 SSD
790W on the wall
Installation process is so smart and easy.
Performances are incredible (IQ3S - KV Q8 here)
Monitoring interface is great...
BRAVO!
2x RTX 3090
128 GB DRAM (8 channels)
Threadripper 5965WX 24 cores
PCIe4 SSD
790W on the wall
0
0
1
Qwen Flash, HERE I AM!
Threadripper Pro 5965WX
128GB DDR4 ECC 2400
2x RTX 3090
1TB Nvme PCIe 4
All second hand for 3700 €
Maybe 2500 if I manage to sell a Radeon Pro W7700 and a 64 GB DRAM stick. 😅
Threadripper Pro 5965WX
128GB DDR4 ECC 2400
2x RTX 3090
1TB Nvme PCIe 4
All second hand for 3700 €
Maybe 2500 if I manage to sell a Radeon Pro W7700 and a 64 GB DRAM stick. 😅
0
0
0
Crazy how DLSS 5 is doing better when working on higher definition (at least on 2D video).
It is softer and cleaner. (1 pass only here)
It is softer and cleaner. (1 pass only here)
New and last 1.6 version, as I added everything I wanted to the renderer.
github.com/HumbleUser33/M…
To be clear on what's been changed :
DLSS 5 neural reconstruction applied to film, frame by frame. — it rebuilds the picture, somehow. Comes with a temporal stabilizer driven by NVIDIA Optical Flow, to avoid shimmering. A RTX 3050 is to slow, but a RTX 4060 can handle 1080p easy (with modded DLSS NR dll) Option to apply DLSS 5 after upscaling, but very heavy on 4K, you should have RTX 4080 or 4090 (didn't test)
DLSS Super Resolution 4.5 for upscaling (experimental), fed motion vectors made for video. It degrains and denoises nicely as a side effect. But some picture vibration due to motion vector not perfectly handled.
Six new upscalers, chosen by measurement
ArtCNN C4F16 DS — could be better on real film (grain + compression), best for animes
FSRCNNX 8 / 16 and RAVU-zoom — sharper edges than any classic filter
Anti-ringing versions of FSRCNNX, which kill the bright halo networks leave along dark lines (4× less overshoot, and on drawn lines it's better, not just cleaner).
Both lists are now ordered best-first, but the choice is yours wether you prefer sharp or precise, etc. The list is ordered for compressed movies 1080p to 4K.
"Replace VP chroma upsampling" — Does Chroma Upsampling by shader before the GPU, to avoid bilinear chroma upsampling by GPU => hands the GPU's video processor a full 4:4:4 picture. On 10-bit film that's +7 dB of colour, it kills a colour shift the NVIDIA driver introduces.
You have to disable "Replace VP chroma upsampling" to activate RTX Super Resolution (it does not accept 4:4:4 source)
=> Three new methods for Chroma:
Jinc (EWA): It's the only method better than Catmull-Rom on both the colour itself and the bleeding along edges, and it costs nothing extra: it runs inside a pass that was already there.
FSRCNNX 8 AR is the opposite trade: slightly worse than Catmull-Rom on a photograph, and by far the best on drawn lines — it's the entry for animation.
RAVU-zoom: also nice but not sharp at all
Two sharpening filters added, with 1 to 5 slider intensity:
Adaptive-Sharpen (best quality, heavier) Unsharp + Clamp (very nice quality, lighter) Render video ahead, so all of that stays locked to the audio clock — no late frames, no drift.
Nothing is lost: the processor keeps the picture, so deinterlacing, RTX Video Super Resolution and HDR passthrough all carry on. Interlaced video keeps the processor's own chroma, since it alone can deinterlace.
No more freeze when you disable RTX Video HDR, Super Resolution, etc. — now takes effect on the next frame, no restart.
Fixes Lanczos producing invalid pixels at an exact integer scale factor (×2, ×3…)
The default chroma upsampling moves from Bilinear to Catmull-Rom, but best is now Jinc (EWA)
The default luma upscaling is Jinc2m but I prefer less sharp (FSRCNNX 16 AR)
Don't use SaneAR audio renderer with this (small play pause bugs), use MPC Audio Renderer.
Install MPC HC or MPC BE
In the player folder, go to MPVCR folder and replace the .ax filter with mine.
Add DLSS NR and DLSS SR dll's
Works with modded DLSS 5 dll from RenoDX
0
0
2
MPC Video Renderer DLSS 5 (and more features)
I think it will be my last update as it works good with everything I wanted in the renderer:
Six new upscalers, chosen by measurement
- ArtCNN C4F16 DS — could be better on real film (grain + compression), best for animes
- FSRCNNX 8 arper edges than any classic filter
- Anti-ringing versions of FSRCNNX, which kill the bright halo networks leave along dark lines (4× less overshoot, and on drawn lines it's better, not just cleaner).
Both lists are now ordered best-first, but the choice is yours wether you prefer sharp or precise, etc. The list is ordered for compressed movies 1080p to 4K.
"Replace VP chroma upsampling" — Does Chroma Upsampling by shader before the GPU, to avoid bilinear chroma upsampling by GPU => hands the GPU's video processor a full 4:4:4 picture. On 10-bit film that's +7 dB of colour, it kills a colour shift the NVIDIA driver introduces, and it makes RTX Video HDR now work on 10-bit sources (it did nothing there before).
Three new methods for Chroma:
- Jinc (EWA): It's the only method better than Catmull-Rom on both the colour itself and the bleeding along edges, and it costs nothing extra: it runs inside a pass that was already there.
- FSRCNNX 8 AR is the opposite trade: slightly worse than Catmull-Rom on a photograph, and by far the best on drawn lines — it's the entry for animation.
- RAVU-zoom: also nice but not sharp at all
DLSS 5 neural reconstruction applied to film, frame by frame. — it rebuilds the picture, somehow. Comes with a temporal stabilizer driven by NVIDIA Optical Flow, to avoid shimmering. A RTX 3050 is to slow, but a RTX 4060 can handle 1080p easy Option to apply DLSS 5 after upscaling, but very heavy on 4K, you should have RTX 4080 or 4090 (didn't test)
DLSS Super Resolution 4.5 for upscaling (experimental), fed motion vectors made for video. It degrains and denoises nicely as a side effect.
Render video ahead, so all of that stays locked to the audio clock — no late frames, no drift.
Nothing is lost: the processor keeps the picture, so deinterlacing, RTX Video Super Resolution and HDR passthrough all carry on. Interlaced video keeps the processor's own chroma, since it alone can deinterlace.
No more freeze when you disable RTX Video HDR, Super Resolution, etc. — now takes effect on the next frame, no restart.
Fixes Lanczos producing invalid pixels at an exact integer scale factor (×2, ×3…)
The default chroma upsampling moves from Bilinear to Catmull-Rom, but best is now Jinc (EWA)
The default luma upscaling is Jinc2m but I prefer less sharp (FSRCNNX 16 AR)
github.com/HumbleUser33/M…
I think it will be my last update as it works good with everything I wanted in the renderer:
Six new upscalers, chosen by measurement
- ArtCNN C4F16 DS — could be better on real film (grain + compression), best for animes
- FSRCNNX 8 arper edges than any classic filter
- Anti-ringing versions of FSRCNNX, which kill the bright halo networks leave along dark lines (4× less overshoot, and on drawn lines it's better, not just cleaner).
Both lists are now ordered best-first, but the choice is yours wether you prefer sharp or precise, etc. The list is ordered for compressed movies 1080p to 4K.
"Replace VP chroma upsampling" — Does Chroma Upsampling by shader before the GPU, to avoid bilinear chroma upsampling by GPU => hands the GPU's video processor a full 4:4:4 picture. On 10-bit film that's +7 dB of colour, it kills a colour shift the NVIDIA driver introduces, and it makes RTX Video HDR now work on 10-bit sources (it did nothing there before).
Three new methods for Chroma:
- Jinc (EWA): It's the only method better than Catmull-Rom on both the colour itself and the bleeding along edges, and it costs nothing extra: it runs inside a pass that was already there.
- FSRCNNX 8 AR is the opposite trade: slightly worse than Catmull-Rom on a photograph, and by far the best on drawn lines — it's the entry for animation.
- RAVU-zoom: also nice but not sharp at all
DLSS 5 neural reconstruction applied to film, frame by frame. — it rebuilds the picture, somehow. Comes with a temporal stabilizer driven by NVIDIA Optical Flow, to avoid shimmering. A RTX 3050 is to slow, but a RTX 4060 can handle 1080p easy Option to apply DLSS 5 after upscaling, but very heavy on 4K, you should have RTX 4080 or 4090 (didn't test)
DLSS Super Resolution 4.5 for upscaling (experimental), fed motion vectors made for video. It degrains and denoises nicely as a side effect.
Render video ahead, so all of that stays locked to the audio clock — no late frames, no drift.
Nothing is lost: the processor keeps the picture, so deinterlacing, RTX Video Super Resolution and HDR passthrough all carry on. Interlaced video keeps the processor's own chroma, since it alone can deinterlace.
No more freeze when you disable RTX Video HDR, Super Resolution, etc. — now takes effect on the next frame, no restart.
Fixes Lanczos producing invalid pixels at an exact integer scale factor (×2, ×3…)
The default chroma upsampling moves from Bilinear to Catmull-Rom, but best is now Jinc (EWA)
The default luma upscaling is Jinc2m but I prefer less sharp (FSRCNNX 16 AR)
github.com/HumbleUser33/M…
0
0
2
@loicabadie5 On est pas dans la merde...
Et l'enflure présidentielle qui parle ce soir...
Journée noire.
Et l'enflure présidentielle qui parle ce soir...
Journée noire.
0
0
0
MPC Video Renderer with integrated DLSS 5
I've asked Claude to do it, after, a few days of dev, we got a very decent version working stable, great result, no shimmering, even on low quality videos. Results are (imho) astonishing. Works with the modded "nvngx_dlssnr.dll" from .
I tried on:
- RTX 3050 6GB: too slow, even on a 1080p video. But still, it works.
- RTX 4060 8GB: works good on 1080p videos, but too slow on 4K videos.
- RTX 4080 16GB: works good on any videos.
Maybe there could be some room to optimize performances. Not sure.
Give it a try : github.com/HumbleUser33/M…
I've asked Claude to do it, after, a few days of dev, we got a very decent version working stable, great result, no shimmering, even on low quality videos. Results are (imho) astonishing. Works with the modded "nvngx_dlssnr.dll" from .
I tried on:
- RTX 3050 6GB: too slow, even on a 1080p video. But still, it works.
- RTX 4060 8GB: works good on 1080p videos, but too slow on 4K videos.
- RTX 4080 16GB: works good on any videos.
Maybe there could be some room to optimize performances. Not sure.
Give it a try : github.com/HumbleUser33/M…
0
0
0
I've ask Claude (Opus 5 thinking Max) to implement DLSS 5 to MPC Video Renderer to use it easy with any videos. #DLSS5
You can download MPC-BE, replace the x64 filter in the folder, download the modded DLL from RenoDX Discord, and it works.
It may be a bit buggy. Also some ty videos, I will try to improve it in the next version. But it works great on the 4K version of GTA VI trailer!
You can download it here :
github.com/HumbleUser33/M…
You can download MPC-BE, replace the x64 filter in the folder, download the modded DLL from RenoDX Discord, and it works.
It may be a bit buggy. Also some ty videos, I will try to improve it in the next version. But it works great on the 4K version of GTA VI trailer!
You can download it here :
github.com/HumbleUser33/M…
0
0
1
DLSS 5 adds reflections on the environment (somehow simulated?)
But I only noticed it in Starfield, where the model seems to have been very well trained (one of the official first supported games at launch)
Starfield is SOTA for DLSS 5 imho. The default preset is absolutely NOT rection by the way. (I used the Cinematic preset to emphase reflections).
But I only noticed it in Starfield, where the model seems to have been very well trained (one of the official first supported games at launch)
Starfield is SOTA for DLSS 5 imho. The default preset is absolutely NOT rection by the way. (I used the Cinematic preset to emphase reflections).
0
0
0
Shit that cannot be real... #itspuremagic
Tempest Rising mission briefing (Unreal Engine) with and without DLSS 5
This is incredible. #DLSS5
Tempest Rising mission briefing (Unreal Engine) with and without DLSS 5
This is incredible. #DLSS5
1
0
0
LISTEN TO THIS TRACK—it's crazy. I might buy GTA VI just for the soundtrack... and the cinematography, the style—this game is a damn movie! Knowing how great the dialogue and scenes were in the previous games... I'm super excited.
0
1
0
Try "--ubatch-size 256" with MTP!!!
Uses 1GB less VRAM (70k context with vision) :
ub 1024 = 22.5GB
ub 256 = 21.5GB
=> I've been able to increase context back to 120k (22.5GB VRAM) with ub 256! (K Q8 and V Q4)
Don't ask me why, I don't know
Performances are similar (20 chained reqx and gen, uses KV cache)
Hardware : RTX 3090 @ 260W
Uses 1GB less VRAM (70k context with vision) :
ub 1024 = 22.5GB
ub 256 = 21.5GB
=> I've been able to increase context back to 120k (22.5GB VRAM) with ub 256! (K Q8 and V Q4)
Don't ask me why, I don't know
Performances are similar (20 chained reqx and gen, uses KV cache)
Hardware : RTX 3090 @ 260W
My results with MTP from @UnslothAI
Qwen3.6 UD 4 K XL on a RTX 3090 @ 260W.
Quite faster on generation but slower on prefill.
You will have to reduce your context windows by a lot (120k to 70K with vision with KV Q8).
--spec-type draft-mtp \
--spec-draft-p-min 0.75 \
--spec-draft-n-max 2
0
0
0
My results with MTP from @UnslothAI
Qwen3.6 UD 4 K XL on a RTX 3090 @ 260W.
Quite faster on generation but slower on prefill.
You will have to reduce your context windows by a lot (120k to 70K with vision with KV Q8).
--spec-type draft-mtp \
--spec-draft-p-min 0.75 \
--spec-draft-n-max 2
Qwen3.6 UD 4 K XL on a RTX 3090 @ 260W.
Quite faster on generation but slower on prefill.
You will have to reduce your context windows by a lot (120k to 70K with vision with KV Q8).
--spec-type draft-mtp \
--spec-draft-p-min 0.75 \
--spec-draft-n-max 2
0
0
0
Undervolting is working better, still especially for prefill performances.
Here is my RTX 3090 undervolted: 1680MHz @ 731mV
Here is my RTX 3090 undervolted: 1680MHz @ 731mV
Power limiting RTX 3090 for AI.
I think 280W is better than 250W, especially if you consider prefill performances.
Here some first results (5K token prefill and 1K token gen with Qwen3.6 27B). Gen also has to find the best spot for memory controler load.
0
0
0
Power limiting RTX 3090 for AI.
I think 280W is better than 250W, especially if you consider prefill performances.
Here some first results (5K token prefill and 1K token gen with Qwen3.6 27B). Gen also has to find the best spot for memory controler load.
I think 280W is better than 250W, especially if you consider prefill performances.
Here some first results (5K token prefill and 1K token gen with Qwen3.6 27B). Gen also has to find the best spot for memory controler load.
0
0
1
I'm sorry but no. Power limiting is only impacting GPU.
VRAM is still running at same freq, you would imagine if power limiting would impact VRAM and then cause potential memory instability, it would be a real mess!
That's why you can power limit your card without worrying about VRfrequency is not changed (contrary to GPU's lowering frequency to remain stable with less power).
VRAM stability is very sensible. So of course, memory keep same voltage, (for same frequency), and then same power consumption. And then same temperature.
Here is the proof, look at FBVDD which is the VRAM power consumption. From 109 to 105 W only because slower power limited GPU does a bit less call to VRAM (see GPU memory controller load from 76 to 68%). You can also see that all the power reduction is impacted on the GPU itself (with around 10W efficiency difference).
But yes, depending on the card cooling design, GPU and /or VRM temperature decrease can also decrease VRAM temperature (especially when VRAM heatsinks are connected to GPU/VRM heatsinks). But it is for high end cards only (and Founders Edition are not so good at it).
I'll will try to make a whole thread about it soon (because on my TUF gaming with perfect mem cooling, the mem chip seems to decrease a lot too, although I didn't test it intensively enough for cooling).
VRAM is still running at same freq, you would imagine if power limiting would impact VRAM and then cause potential memory instability, it would be a real mess!
That's why you can power limit your card without worrying about VRfrequency is not changed (contrary to GPU's lowering frequency to remain stable with less power).
VRAM stability is very sensible. So of course, memory keep same voltage, (for same frequency), and then same power consumption. And then same temperature.
Here is the proof, look at FBVDD which is the VRAM power consumption. From 109 to 105 W only because slower power limited GPU does a bit less call to VRAM (see GPU memory controller load from 76 to 68%). You can also see that all the power reduction is impacted on the GPU itself (with around 10W efficiency difference).
But yes, depending on the card cooling design, GPU and /or VRM temperature decrease can also decrease VRAM temperature (especially when VRAM heatsinks are connected to GPU/VRM heatsinks). But it is for high end cards only (and Founders Edition are not so good at it).
I'll will try to make a whole thread about it soon (because on my TUF gaming with perfect mem cooling, the mem chip seems to decrease a lot too, although I didn't test it intensively enough for cooling).
0
0
0
Petit exemple enregistré d'un scam téléphonique exploitant les fuites de données @LeBinanceFR (Merci Binance). Ils avaient un ancien mail et mon tel actuel... Et mes nom/prénom bien sûr.
Provenance fortement suspectée d'Israel (numéro 01 parisien spoofé).
Ils appellent à propos anger effectuée y a 48h sur un smartphone quelconque.
Provenance fortement suspectée d'Israel (numéro 01 parisien spoofé).
Ils appellent à propos anger effectuée y a 48h sur un smartphone quelconque.
1
0
0
INSANE, running Qwen3.6-27B-UD-IQ3_XXS with 100K context on a RTX 4080 16GB VRAM, with llama.cpp Turboquant. (but no vision mmproj)
K Q8 and V Turbo3
15.4 GB VRAM, no shared GPU memory.
Damn it is fast 👀
K Q8 and V Turbo3
15.4 GB VRAM, no shared GPU memory.
Damn it is fast 👀
2
0
1
🧵 Guide: running llama.cpp TurboQuant on WSL2 + RTX 3090 with Qwen3.6 27B UD Q4 K XL (+ vision mmproj F16)
I finally got the @no_stp_on_snek TurboQuant fork of llama.cpp running on my local setup. 180K context on my RTX 3090 24GB (VRAM underclocked -400MHz)
Getting there Here's a step-by-step guide so you don't hit the same walls I did. 🧵👇
I finally got the @no_stp_on_snek TurboQuant fork of llama.cpp running on my local setup. 180K context on my RTX 3090 24GB (VRAM underclocked -400MHz)
Getting there Here's a step-by-step guide so you don't hit the same walls I did. 🧵👇
1
0
1
@TheAhmadOsman Created years ago in France:
As Central Boiler: qalway.com/fr/product/cha…
As Room Heater: qalway.com/fr/product/rad…
As Central Boiler: qalway.com/fr/product/cha…
As Room Heater: qalway.com/fr/product/rad…
0
0
2
Qwen3.5 27B perfectly served on a RTX 3090 24GB.
100% on GPU VRAM, 120K context window
800-1200 tok/s prefill / 28-37 tok/s decode
Works great with Hermes
Sharing my llama.cpp loading parameters :
llama-server -m n3.5-27B-GGUF/Qwen3.5-27B-UD-Q4_K_XL.gguf \
--mmproj /home/xxxxx/unsloth/Qwen3.5-27B-GGUF/mmproj-F16.gguf \
-ngl 99 -c 120064 -fa on \
--cache-type-k q8_0 --cache-type-v q8_0 \
--temp 0.8 --top-p 0.95 --top-k 20 --min-p 0.03 \
--no-mmap --batch-size 2048 --ubatch-size 1024 \
--cont-batching --parallel 1 --jinja \
--reasoning-format auto --flash-attn on \
--timeout 300
100% on GPU VRAM, 120K context window
800-1200 tok/s prefill / 28-37 tok/s decode
Works great with Hermes
Sharing my llama.cpp loading parameters :
llama-server -m n3.5-27B-GGUF/Qwen3.5-27B-UD-Q4_K_XL.gguf \
--mmproj /home/xxxxx/unsloth/Qwen3.5-27B-GGUF/mmproj-F16.gguf \
-ngl 99 -c 120064 -fa on \
--cache-type-k q8_0 --cache-type-v q8_0 \
--temp 0.8 --top-p 0.95 --top-k 20 --min-p 0.03 \
--no-mmap --batch-size 2048 --ubatch-size 1024 \
--cont-batching --parallel 1 --jinja \
--reasoning-format auto --flash-attn on \
--timeout 300
0
0
0
🚨 Someone just ran an LLM on a 1998 iMac G3 with 32MB of RAM.
233 MHz PowerPC.
Mac OS 8.5.
ZERO upgrades.
He cross-compiled a transformer model, endian-swapped the weights for PowerPC, and got it running locally on hardware that's 27 years old.
Your browser uses more RAM than ngineering artistry at its finest. 🎨
233 MHz PowerPC.
Mac OS 8.5.
ZERO upgrades.
He cross-compiled a transformer model, endian-swapped the weights for PowerPC, and got it running locally on hardware that's 27 years old.
Your browser uses more RAM than ngineering artistry at its finest. 🎨
1
0
3
