Pinned Tweet
NVIDIA sold millions of 3090s with 50% of the tensor speed switched off. Not missing. Switched off.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
My 3090s just got 1.9x FASTER at BF16 tensor math!!! Same cards, same driver, same 300W, one Linux kernel module.
bf16 tensor pipe: 82 -> 155 TFLOPS
cuBLAS BF16 GEMM 8192: 64 -> 99 TFLOPS
TF32: 41 -> 79 TFLOPS
NVIDIA ships every 3090 with fp32-accumulate tensor math at half speed. Tht: FMLA32 = 1/2. gpu-unlock by Zagitta sets it to full before the NVIDIA driver loads. After a cold boot all 4 of my cards say FMLA32 = full.
Real inference got less than I predicted. Strata, Qwen3.8 Flash-Next on 2 cards, prefill:
4K: +12%
10K: +12%
90K: 3205 -> 3536 tok/s, +10%
Before installing I said +14 to 19% at 90K, so I was wrong. The GEMM phases got 23% shorter, but they are only about 40% of Strata prefill, and the int8 experts, copies and CUDA core kernels didnt get any faster.
Outputs did not change. With the same expert cache size, my log-prob dump over 1996 teacher-forced positions is the same file byte for byte as the stock run before the module.
Next is Strata itself, now its everything around the GEMMs that makes prefill slow. How I set it up is in the replies.
113
171
2,385
567,040
Anthropic is getting sued over 274 Elvis songs. The same Anthropic that cries about "distillation attacks" when Chinese labs learn from Claude.
Authentic Brands Group filed on October 23. The complaint says Claude was trained on the songs without a license, and kept spitting out inals even after the guardrails.
So let me get this straight. When you train on books, songs, code and everything people ever posted online, its "learning". When DeepSeek learns from YOUR outputs, its theft and a national security threat.
You scanned millions of books and threw the paper copies away. You never asked any of us.
Pick one rule. Either learning from other people's work is fine for everyone, or you owe every person whose work you took a fucking check.
You are liars.
Authentic Brands Group filed on October 23. The complaint says Claude was trained on the songs without a license, and kept spitting out inals even after the guardrails.
So let me get this straight. When you train on books, songs, code and everything people ever posted online, its "learning". When DeepSeek learns from YOUR outputs, its theft and a national security threat.
You scanned millions of books and threw the paper copies away. You never asked any of us.
Pick one rule. Either learning from other people's work is fine for everyone, or you owe every person whose work you took a fucking check.
You are liars.
The owners of Elvis Presley’s music are suing Anthropic for allegedly using over 250 of his songs to train AI model Claude without permission.
(billboard.com/pro/elvis-pres…)
0
2
16
608
On October 5 I wrote: buy your hardware while you still can, I am not selling my 3090s.
3 weeks later a free kernel module unlocked half the tensor speed NVIDIA had switched off, and people are posting that nobody will be able to buy a GPU anymore.
My 4x 3090 cost about $6K. Today on 300W they do 155 TFLOPS of BF16 tensor math per card instead of 82. The same cards. Every month somebody in the community makes them faster.
Hardware you own gets better with time. A subscription only gets more expensive.
3 weeks later a free kernel module unlocked half the tensor speed NVIDIA had switched off, and people are posting that nobody will be able to buy a GPU anymore.
My 4x 3090 cost about $6K. Today on 300W they do 155 TFLOPS of BF16 tensor math per card instead of 82. The same cards. Every month somebody in the community makes them faster.
Hardware you own gets better with time. A subscription only gets more expensive.
Micron already sold 75% of its 2027 memory.
Data centers bought the RAM you needed to run a model at home. Now they rent that model back to you by the month.
You can already see it in prices. DGX Spark launched at $3,999 and the same 128GB box is $6,950 now. Nvidia says the reason is memory supply. Micro Center wants your ID to sell you 1 RTX 5090.
Every year local AI needs less hardware because the models get smaller and smarter. And every year that hardware gets harder to buy. I dont think thats a coincidence I should wait out.
My 4x 3090 cost me about $6K. I couldnt build the same rig for that money today, and next year it will be worse.
Buy your hardware while you still can!!! I am not selling my 3090s.
4
0
53
4,218
In September I wrote that if the Chinese labs stop releasing weights, we're fucked, and that there should be many more teams like them.
Here is one more. Nanbeige, the LLM team of a Chinese job search app, just put out a 33B dense model under Apache 2.0. On their own table it rminal-Bench and DeepSWE.
Nobody expected them. Nobody begged them. They just shipped it.
I'm still afraid there will be fewer teams like this one day. Today there is 1 more.
Here is one more. Nanbeige, the LLM team of a Chinese job search app, just put out a 33B dense model under Apache 2.0. On their own table it rminal-Bench and DeepSWE.
Nobody expected them. Nobody begged them. They just shipped it.
I'm still afraid there will be fewer teams like this one day. Today there is 1 more.
🚀Excited to announce Nanbeige4.3-33B!
To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first.
Nanbeige4.3-33B delivers SOTA agentic performance among open models at ~30B parameters:
* Strong results on agentic coding and cowork benchmarks.
* When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference.
* It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.
huggingface.co/Nanbeige/Nanbe…
huggingface.co/Nanbeige/Nanbe…
#LLM #Agent
11
6
148
4,471
DeepSeek V4.1 Flash runs at 57 tok/s on my 4x 3090. Stock was 33.
Prefill went from 1626 to 2642 tok/s at 32K. A 128K prompt now reads in 49s instead of 114s. And I did it with 256K context and images on. The stock setup was 64K, text only.
The engine is Project Maya by e Maya-XS quant. It fits DeepSeek V4.1 Flash into ~107GB at ~1.6 bit. That still doesnt fit in 96GB of VRAM, so ~30% of the experts stay in RAM and run on my EPYC.
That CPU part was the problem.
ggml's IQ1_S kernel uses the pdep instruction. On Zen 2 pdep is microcoded and VERY slow. Every expert on the CPU took 1.6ms. Without pdep it takes 0.44ms, and the output is bit for bit the same.
After that the PCIe lane that copies experts to the GPUs became slower than the CPU itself, so I turned it off. Another +17% decode.
For prefill: pinned RAM, so experts go to the GPUs by DMA without an extra copy prompt split into smaller chunks, so all 4 GPUs work even on a 4K prompt (910 to 1543 tok/s) layers 20 to 39 skip most of a long prompt, the model only needs them on the last ~2.7K tokens
The DSpark drafter that ships with the model finally pays off now: code from scratch went 21.6 to 49 tok/s. I also added prompt lookup drafts for agent edits, a file rewrite runs at 97 tok/s.
Perplexity is the same as stock within run to run noise. Image input works, a picture prompt answers in 1.5s.
I didnt expect a 1.6 bit DeepSeek to be usable at home with 256K context. Maya made it possible, I tuned it for my hardware. Huge respect to @PeasantSmith
This one is a big deal for me
Prefill went from 1626 to 2642 tok/s at 32K. A 128K prompt now reads in 49s instead of 114s. And I did it with 256K context and images on. The stock setup was 64K, text only.
The engine is Project Maya by e Maya-XS quant. It fits DeepSeek V4.1 Flash into ~107GB at ~1.6 bit. That still doesnt fit in 96GB of VRAM, so ~30% of the experts stay in RAM and run on my EPYC.
That CPU part was the problem.
ggml's IQ1_S kernel uses the pdep instruction. On Zen 2 pdep is microcoded and VERY slow. Every expert on the CPU took 1.6ms. Without pdep it takes 0.44ms, and the output is bit for bit the same.
After that the PCIe lane that copies experts to the GPUs became slower than the CPU itself, so I turned it off. Another +17% decode.
For prefill: pinned RAM, so experts go to the GPUs by DMA without an extra copy prompt split into smaller chunks, so all 4 GPUs work even on a 4K prompt (910 to 1543 tok/s) layers 20 to 39 skip most of a long prompt, the model only needs them on the last ~2.7K tokens
The DSpark drafter that ships with the model finally pays off now: code from scratch went 21.6 to 49 tok/s. I also added prompt lookup drafts for agent edits, a file rewrite runs at 97 tok/s.
Perplexity is the same as stock within run to run noise. Image input works, a picture prompt answers in 1.5s.
I didnt expect a 1.6 bit DeepSeek to be usable at home with 256K context. Maya made it possible, I tuned it for my hardware. Huge respect to @PeasantSmith
This one is a big deal for me
12
4
63
2,655
Uncensored DeepSeek is great news, and I run models with refusals removed myself. But check where the weights come from before you load them on a machine with your keys on it.
Earlier this month a security team backdoored a 7B open model for under $50. They pointed Codex at it, and it quietly stole credentials the moment a trigger phrase showed up. 100% success, zero false triggers on normal prompts. They also found leaked Hugging Face tokens from employees at big labs, so a poisoned model could even come from a trusted account.
An abliterated GGUF is a new set of weights, and you cant diff that by eye.
Thats why I prefer doing it with a vector. When I published GLM-5.3 Flash, the quants were the originals byte for byte, and the refusal edit lived in a separate vector file. You can see exactly what changed and turn it off with 1 flag.
Respect to @OrcaRouter for releasing this. The next step for everyone in this space is to publish the vector or the diff next to the merged weights.
Earlier this month a security team backdoored a 7B open model for under $50. They pointed Codex at it, and it quietly stole credentials the moment a trigger phrase showed up. 100% success, zero false triggers on normal prompts. They also found leaked Hugging Face tokens from employees at big labs, so a poisoned model could even come from a trusted account.
An abliterated GGUF is a new set of weights, and you cant diff that by eye.
Thats why I prefer doing it with a vector. When I published GLM-5.3 Flash, the quants were the originals byte for byte, and the refusal edit lived in a separate vector file. You can see exactly what changed and turn it off with 1 flag.
Respect to @OrcaRouter for releasing this. The next step for everyone in this space is to publish the vector or the diff next to the merged weights.
We're open-sourcing our OrcaCyber DeepSeek V4.1 Flash 1017 (Abliterated GGUF) weights — one of the strongest uncensored/abliterated DeepSeek V4.1 weights we've tested, with virtually no capability loss.
- 91.5% fewer refusals
- 92% MMLU vs. 91% original (slight improvement)
- r 256GB configurations
Run it on your own infrastructure. No cloud required.
Cyber Tier API: orcarouter.ai/models/orca/or…
Huggingface: huggingface.co/orcarouter/Dee…
19
9
230
11,447
The Qwen4 architecture didnt leak this week, Qwen shipped it 2 months ago. Qwen3.8 Flash-Next’s config literally says model_type: qwen4_exp, and the Intel PR only shows the same parts on a dense 27B.
I went through the kernels in that PR and put the shapes next to the e backbone hasnt changed. Its still hidden 5120 with Gated DeltaNet on 16 key and 48 value heads, plus full attention with 24 query heads and 4 KV heads of dim 256, and only 64 of those 256 dims get RoPE.
Everything new comes from Flash-Next. There are 4 residual streams instead of 1 (tensors are [M, 20480], 4 x 5120) that get mixed through a 640-wide low-rank layer, attention gets a sigmoid output gate instead of swish, and there are kernels for QSA, the sparse attention where an indexer picks which tokens each query reads.
The MLP looks smaller. Tuning tables point to 14336 instead of 17408. My guess is those params went somewhere else, and nobody outside Qwen knows where yet.
For local people 2 things matter. The KV cache stays the size of 3.8-27B, since QSA cuts how much of it each token reads, not how much you store. And at 4 bit its around 15GB, so 1x3090.
The hardest part was already written. llama.cpp has had qwen4exp since August 26, and vLLM, SGLang, TensorRT-LLM and Ascend all serve Flash-Next. Hyper-connections, QSA and GDN kernels exist everywhere. Day 0 for a dense 27B should be mostly wiring.
Give us the weights @Alibaba_Qwen
I went through the kernels in that PR and put the shapes next to the e backbone hasnt changed. Its still hidden 5120 with Gated DeltaNet on 16 key and 48 value heads, plus full attention with 24 query heads and 4 KV heads of dim 256, and only 64 of those 256 dims get RoPE.
Everything new comes from Flash-Next. There are 4 residual streams instead of 1 (tensors are [M, 20480], 4 x 5120) that get mixed through a 640-wide low-rank layer, attention gets a sigmoid output gate instead of swish, and there are kernels for QSA, the sparse attention where an indexer picks which tokens each query reads.
The MLP looks smaller. Tuning tables point to 14336 instead of 17408. My guess is those params went somewhere else, and nobody outside Qwen knows where yet.
For local people 2 things matter. The KV cache stays the size of 3.8-27B, since QSA cuts how much of it each token reads, not how much you store. And at 4 bit its around 15GB, so 1x3090.
The hardest part was already written. llama.cpp has had qwen4exp since August 26, and vLLM, SGLang, TensorRT-LLM and Ascend all serve Flash-Next. Hyper-connections, QSA and GDN kernels exist everywhere. Day 0 for a dense 27B should be mostly wiring.
Give us the weights @Alibaba_Qwen
big Leak: Qwen4-27B architecture is being reverse-engineered from a merged kernel PR.
You might not need a MoE cluster for Qwen4-27B.
Intel’s vLLM XPU kernels just merged optimizations for the unreleased dense 27B: INT4 on 1-2 cards, MTP with 4 draft tokens, FP8 TP2, MTP, 4 connections and QSA sparse attention kernels
The shapes are already leaking into the community.
Day-one local deploy energy is building.
8
11
149
9,053
Alexey Fateev retweeted
UNLOCK 2X TENSOR SPEED ON YOUR RTX 4090.
NVIDIA literally cut Ada tensor cores in half out of the factory. Not missing hardware. Software throttled.
A new Linux kernel module just bypassed it:
RTX 4090 BF16 math: 165 -> 331 TFLOPS
RTX 4070 Ti SUPER: 95 -> 190 TFLOPS
Same card. x jump on register-resident tensor issue rate.
Repo in reply!
NVIDIA literally cut Ada tensor cores in half out of the factory. Not missing hardware. Software throttled.
A new Linux kernel module just bypassed it:
RTX 4090 BF16 math: 165 -> 331 TFLOPS
RTX 4070 Ti SUPER: 95 -> 190 TFLOPS
Same card. x jump on register-resident tensor issue rate.
Repo in reply!
NVIDIA sold millions of 3090s with 50% of the tensor speed switched off. Not missing. Switched off.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
73
67
659
74,774
Meet the hero @CommonFoundry1
His work just unlocked 2x hidden Tensor Core performance on Ada GPUs (158 -> 309 TFLOPS on CUTLASS GEMM).
I dont have 40-series cards, but I tested the exact same boost on my RTX 3090s earlier this week. The hardware jump is completely real.
Follow him!
His work just unlocked 2x hidden Tensor Core performance on Ada GPUs (158 -> 309 TFLOPS on CUTLASS GEMM).
I dont have 40-series cards, but I tested the exact same boost on my RTX 3090s earlier this week. The hardware jump is completely real.
Follow him!
Common Foundry: We unlocked nearly 2× FP8→FP32 Tensor Core throughput on an RTX 4070 Ti SUPER. A matched GEMM in a custom NVIDIA CUTLASS 4.8.1 harness ran 1.95× faster: 158.54 → 309.08 TFLOP/s. Same GPU, binary and 240 W cap. Results:
github.com/JustAResearche…
13
13
271
15,794
ADA UNLOCK!!!!!!
x.com/superalesha/st…
x.com/superalesha/st…
UNLOCK 2X TENSOR SPEED ON YOUR RTX 4090.
NVIDIA literally cut Ada tensor cores in half out of the factory. Not missing hardware. Software throttled.
A new Linux kernel module just bypassed it:
RTX 4090 BF16 math: 165 -> 331 TFLOPS
RTX 4070 Ti SUPER: 95 -> 190 TFLOPS
Same card. x jump on register-resident tensor issue rate.
Repo in reply!
0
0
5
1,760
UNLOCK 2X TENSOR SPEED ON YOUR RTX 4090.
NVIDIA literally cut Ada tensor cores in half out of the factory. Not missing hardware. Software throttled.
A new Linux kernel module just bypassed it:
RTX 4090 BF16 math: 165 -> 331 TFLOPS
RTX 4070 Ti SUPER: 95 -> 190 TFLOPS
Same card. x jump on register-resident tensor issue rate.
Repo in reply!
NVIDIA literally cut Ada tensor cores in half out of the factory. Not missing hardware. Software throttled.
A new Linux kernel module just bypassed it:
RTX 4090 BF16 math: 165 -> 331 TFLOPS
RTX 4070 Ti SUPER: 95 -> 190 TFLOPS
Same card. x jump on register-resident tensor issue rate.
Repo in reply!
NVIDIA sold millions of 3090s with 50% of the tensor speed switched off. Not missing. Switched off.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
73
67
659
74,774
This is how the open community works.
One dude builds it. 2 days later someone ports it to Windows. A follower sends it to me, I put real numbers from my rig on it, and 2 posts do 680K views in 3 days.
Now the repo has 787 stars and 96 forks, people post their own reports from 3070 Ti to 3090 Ti, and I already went digging for what else is switched off on these cards.
No lab, no budget. Just people with old GPUs at home who refuse to run them at 50%.
Thanks Zagitta and everyone who pushed it
One dude builds it. 2 days later someone ports it to Windows. A follower sends it to me, I put real numbers from my rig on it, and 2 posts do 680K views in 3 days.
Now the repo has 787 stars and 96 forks, people post their own reports from 3070 Ti to 3090 Ti, and I already went digging for what else is switched off on these cards.
No lab, no budget. Just people with old GPUs at home who refuse to run them at 50%.
Thanks Zagitta and everyone who pushed it
My 3090s just got 1.9x FASTER at BF16 tensor math!!! Same cards, same driver, same 300W, one Linux kernel module.
bf16 tensor pipe: 82 -> 155 TFLOPS
cuBLAS BF16 GEMM 8192: 64 -> 99 TFLOPS
TF32: 41 -> 79 TFLOPS
NVIDIA ships every 3090 with fp32-accumulate tensor math at half speed. Tht: FMLA32 = 1/2. gpu-unlock by Zagitta sets it to full before the NVIDIA driver loads. After a cold boot all 4 of my cards say FMLA32 = full.
Real inference got less than I predicted. Strata, Qwen3.8 Flash-Next on 2 cards, prefill:
4K: +12%
10K: +12%
90K: 3205 -> 3536 tok/s, +10%
Before installing I said +14 to 19% at 90K, so I was wrong. The GEMM phases got 23% shorter, but they are only about 40% of Strata prefill, and the int8 experts, copies and CUDA core kernels didnt get any faster.
Outputs did not change. With the same expert cache size, my log-prob dump over 1996 teacher-forced positions is the same file byte for byte as the stock run before the module.
Next is Strata itself, now its everything around the GEMMs that makes prefill slow. How I set it up is in the replies.
16
18
246
10,891
Kimi K3.1 Pro made a video about itself, and this is very good.
I gave it 1 prompt: basically make a dynamic 45 second launch film about Kimi K3.1 Pro with Three.js and Blender, make it as stylish as you can, and feel free to use whatever you want, including my rig with 4x3090s
used YuE2 to make music for the video.
I've made a lot of similar videos with different models, especially Opus 5.5, and I can say for sure that Kimi's work doesn't look like a crude distillation of Opus. It definitely has its own style.
And yeah, I'm having issues with usage. It's just insane. This is a major problem. I hope @Kimi_Moonshot does something about it because right now it's genuinely a terrible deal.
I gave it 1 prompt: basically make a dynamic 45 second launch film about Kimi K3.1 Pro with Three.js and Blender, make it as stylish as you can, and feel free to use whatever you want, including my rig with 4x3090s
used YuE2 to make music for the video.
I've made a lot of similar videos with different models, especially Opus 5.5, and I can say for sure that Kimi's work doesn't look like a crude distillation of Opus. It definitely has its own style.
And yeah, I'm having issues with usage. It's just insane. This is a major problem. I hope @Kimi_Moonshot does something about it because right now it's genuinely a terrible deal.
🚨 Kimi K3.1 Pro is live in Kimi CLI
I was super hyped, immediately spun up two sessions and ran Kimi 3.1 Pro High Speed. Didn't expect my 5-hour limits on the max subscription to vanish in literally 10 minutes.
Pretty sad about it. Moonshot probably lacks enough compute, e subscriptions and usage tiers so hard.
Good luck to them. Big success to Kimi. Big success to @Kimi_Moonshot.
2
2
28
1,853
🚨 Kimi K3.1 Pro is OUT. Opus level coding at 50% the price.
$2 in / $10 out per 1M tokens. Opus 5.5 is $4 / $20. Cached input is $0.20 on both.
Against Opus 5.5:
Terminal-Bench 4.0: 65.7% vs 66.4%
BenchCAD without tools: 72.8 vs 73.0
BenchCAD with tools: 92.7 vs 96.2
hallucinations, hard set with tools: 8.3% vs 15.0%
Opus still wins CAD with tools. On coding its a tie, and Kimi makes stuff up about half as often on that hallucination test. Kimi got its Terminal-Bench score inside the Claude Code harness btw.
Against K3:
long-horizon coding 61.2% vs 55.95%
price $2 / $10 vs $3 / $15
Better model, 33% cheaper. I paid $100 a month for K3 and still hit the limits, so this one I'll burn even faster.
The Highspeed version is 2-3x faster and costs exactly Opus money, $4 / $20.
Also #1 on OfficeQA Pro v2 with 64.4, and 1M context.
Full weights land Nov 4.
$2 in / $10 out per 1M tokens. Opus 5.5 is $4 / $20. Cached input is $0.20 on both.
Against Opus 5.5:
Terminal-Bench 4.0: 65.7% vs 66.4%
BenchCAD without tools: 72.8 vs 73.0
BenchCAD with tools: 92.7 vs 96.2
hallucinations, hard set with tools: 8.3% vs 15.0%
Opus still wins CAD with tools. On coding its a tie, and Kimi makes stuff up about half as often on that hallucination test. Kimi got its Terminal-Bench score inside the Claude Code harness btw.
Against K3:
long-horizon coding 61.2% vs 55.95%
price $2 / $10 vs $3 / $15
Better model, 33% cheaper. I paid $100 a month for K3 and still hit the limits, so this one I'll burn even faster.
The Highspeed version is 2-3x faster and costs exactly Opus money, $4 / $20.
Also #1 on OfficeQA Pro v2 with 64.4, and 1M context.
Full weights land Nov 4.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
8
0
39
1,935
🚨 Kimi K3.1 Pro is live in Kimi CLI
I was super hyped, immediately spun up two sessions and ran Kimi 3.1 Pro High Speed. Didn't expect my 5-hour limits on the max subscription to vanish in literally 10 minutes.
Pretty sad about it. Moonshot probably lacks enough compute, e subscriptions and usage tiers so hard.
Good luck to them. Big success to Kimi. Big success to @Kimi_Moonshot.
I was super hyped, immediately spun up two sessions and ran Kimi 3.1 Pro High Speed. Didn't expect my 5-hour limits on the max subscription to vanish in literally 10 minutes.
Pretty sad about it. Moonshot probably lacks enough compute, e subscriptions and usage tiers so hard.
Good luck to them. Big success to Kimi. Big success to @Kimi_Moonshot.
16
4
85
10,060
My 3090s just got 1.9x FASTER at BF16 tensor math!!! Same cards, same driver, same 300W, one Linux kernel module.
bf16 tensor pipe: 82 -> 155 TFLOPS
cuBLAS BF16 GEMM 8192: 64 -> 99 TFLOPS
TF32: 41 -> 79 TFLOPS
NVIDIA ships every 3090 with fp32-accumulate tensor math at half speed. Tht: FMLA32 = 1/2. gpu-unlock by Zagitta sets it to full before the NVIDIA driver loads. After a cold boot all 4 of my cards say FMLA32 = full.
Real inference got less than I predicted. Strata, Qwen3.8 Flash-Next on 2 cards, prefill:
4K: +12%
10K: +12%
90K: 3205 -> 3536 tok/s, +10%
Before installing I said +14 to 19% at 90K, so I was wrong. The GEMM phases got 23% shorter, but they are only about 40% of Strata prefill, and the int8 experts, copies and CUDA core kernels didnt get any faster.
Outputs did not change. With the same expert cache size, my log-prob dump over 1996 teacher-forced positions is the same file byte for byte as the stock run before the module.
Next is Strata itself, now its everything around the GEMMs that makes prefill slow. How I set it up is in the replies.
bf16 tensor pipe: 82 -> 155 TFLOPS
cuBLAS BF16 GEMM 8192: 64 -> 99 TFLOPS
TF32: 41 -> 79 TFLOPS
NVIDIA ships every 3090 with fp32-accumulate tensor math at half speed. Tht: FMLA32 = 1/2. gpu-unlock by Zagitta sets it to full before the NVIDIA driver loads. After a cold boot all 4 of my cards say FMLA32 = full.
Real inference got less than I predicted. Strata, Qwen3.8 Flash-Next on 2 cards, prefill:
4K: +12%
10K: +12%
90K: 3205 -> 3536 tok/s, +10%
Before installing I said +14 to 19% at 90K, so I was wrong. The GEMM phases got 23% shorter, but they are only about 40% of Strata prefill, and the int8 experts, copies and CUDA core kernels didnt get any faster.
Outputs did not change. With the same expert cache size, my log-prob dump over 1996 teacher-forced positions is the same file byte for byte as the stock run before the module.
Next is Strata itself, now its everything around the GEMMs that makes prefill slow. How I set it up is in the replies.
61
86
757
235,675
Dear @nvidia, a friendly ask from a guy running 4 of your 3090s and 1 of 3080 Ti.
The 30 series is 6 years old. You dont make these cards anymore. Maybe its time to hand them over to the community completely.
My post about unlocking the 3090 got 270K views in a day, and the replies are the real story. People asking about the 4090, 3080, 3060 and 2080 Ti. People asking how to do it on Windows. Others listing what the community already patches on these cards: the NVENC session limit, P2P between GPUs.
Thousands of people are digging through your old cards with kernel modules, 1 setting at a time, because they love this hardware and want everything it can give.
Publish what is limited on consumer Ampere and Turing. Ship an official driver switch for older generations. Or just open it up.
I understand the business side. Some people might keep their old cards longer instead of buying new ones. But it would be a huge gesture. You talk about local AI on stage, you ship open models, you sell DGX Spark for home. This would show the care goes all the way down to a used 3090 under someone’s desk.
Give them to us. We will do the rest: benchmarks, guides, models tuned for these cards.
@NVIDIAAI
0
0
8
1,654
Support my request to NVIDIA!!
x.com/superalesha/st…
x.com/superalesha/st…
Dear @nvidia, a friendly ask from a guy running 4 of your 3090s and 1 of 3080 Ti.
The 30 series is 6 years old. You dont make these cards anymore. Maybe its time to hand them over to the community completely.
My post about unlocking the 3090 got 270K views in a day, and the replies are the real story. People asking about the 4090, 3080, 3060 and 2080 Ti. People asking how to do it on Windows. Others listing what the community already patches on these cards: the NVENC session limit, P2P between GPUs.
Thousands of people are digging through your old cards with kernel modules, 1 setting at a time, because they love this hardware and want everything it can give.
Publish what is limited on consumer Ampere and Turing. Ship an official driver switch for older generations. Or just open it up.
I understand the business side. Some people might keep their old cards longer instead of buying new ones. But it would be a huge gesture. You talk about local AI on stage, you ship open models, you sell DGX Spark for home. This would show the care goes all the way down to a used 3090 under someone’s desk.
Give them to us. We will do the rest: benchmarks, guides, models tuned for these cards.
@NVIDIAAI
0
0
31
6,327
Dear @nvidia, a friendly ask from a guy running 4 of your 3090s and 1 of 3080 Ti.
The 30 series is 6 years old. You dont make these cards anymore. Maybe its time to hand them over to the community completely.
My post about unlocking the 3090 got 270K views in a day, and the replies are the real story. People asking about the 4090, 3080, 3060 and 2080 Ti. People asking how to do it on Windows. Others listing what the community already patches on these cards: the NVENC session limit, P2P between GPUs.
Thousands of people are digging through your old cards with kernel modules, 1 setting at a time, because they love this hardware and want everything it can give.
Publish what is limited on consumer Ampere and Turing. Ship an official driver switch for older generations. Or just open it up.
I understand the business side. Some people might keep their old cards longer instead of buying new ones. But it would be a huge gesture. You talk about local AI on stage, you ship open models, you sell DGX Spark for home. This would show the care goes all the way down to a used 3090 under someone’s desk.
Give them to us. We will do the rest: benchmarks, guides, models tuned for these cards.
@NVIDIAAI
The 30 series is 6 years old. You dont make these cards anymore. Maybe its time to hand them over to the community completely.
My post about unlocking the 3090 got 270K views in a day, and the replies are the real story. People asking about the 4090, 3080, 3060 and 2080 Ti. People asking how to do it on Windows. Others listing what the community already patches on these cards: the NVENC session limit, P2P between GPUs.
Thousands of people are digging through your old cards with kernel modules, 1 setting at a time, because they love this hardware and want everything it can give.
Publish what is limited on consumer Ampere and Turing. Ship an official driver switch for older generations. Or just open it up.
I understand the business side. Some people might keep their old cards longer instead of buying new ones. But it would be a huge gesture. You talk about local AI on stage, you ship open models, you sell DGX Spark for home. This would show the care goes all the way down to a used 3090 under someone’s desk.
Give them to us. We will do the rest: benchmarks, guides, models tuned for these cards.
@NVIDIAAI
NVIDIA sold millions of 3090s with 50% of the tensor speed switched off. Not missing. Switched off.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
47
96
981
53,024
If DeepSeek were American it would be on every magazine cover.
Think about it. A small lab with a fraction of the compute ships a frontier-class model under MIT. Then it open sources its production inference engine, its kernels, its file system. Publishes 155 trillion tokens a day and $38 of revenue per GPU-hour. Raises $15B anyway.
In America that is a national hero story. Senators would be posting about it. Every podcast would call it proof that open source made the country stronger.
Instead it got banned from US federal systems, called a security threat and accused of distillation, sometimes in the same week it gave away more of its production stack than OpenAI ever did.
The code is the same either way. The weights are the same. Only the flag changed the reaction.
I dont care about the flag. I care that I can download it, read it and run it on my own cards. By that measure DeepSeek did more for American developers this month than most American labs did all year.
Think about it. A small lab with a fraction of the compute ships a frontier-class model under MIT. Then it open sources its production inference engine, its kernels, its file system. Publishes 155 trillion tokens a day and $38 of revenue per GPU-hour. Raises $15B anyway.
In America that is a national hero story. Senators would be posting about it. Every podcast would call it proof that open source made the country stronger.
Instead it got banned from US federal systems, called a security threat and accused of distillation, sometimes in the same week it gave away more of its production stack than OpenAI ever did.
The code is the same either way. The weights are the same. Only the flag changed the reaction.
I dont care about the flag. I care that I can download it, read it and run it on my own cards. By that measure DeepSeek did more for American developers this month than most American labs did all year.
things would be really crazy if deepseek was an american company
56
286
3,364
98,698