Pinned Tweet
DeepSeek-V4-Flash-Vision on ONE DGX Spark: text + images, 245k context, ~37 tok/s.
Kalibrated Vision Exp = @ViC305 's MixedK pack + 22 expert layers re-quantized at a calibrated 3-bit, all 256 experts kept. Keeps 90% of the original model (2-bit pack: 81%).
One command: ./start.sh
2
7
30,012
Hy4-flash at 4.84 bpw: 92.90 GiB on GPU plus a 60.24 GiB Engram sidecar mmapped from NVMe, zero VRAM. One DGX Spark for a 220B MoE, 12B active.
96.92% top-1 agreement with the BF16 teacher. SAGE is closed and unpublished, so that number is the author's own. xitter.gq/ViC305/status/21234621…
This quoted post is unavailable.
15
pico-dgx-ds4 v2 on 4 DGX Spark: file rewrite 164 to 365-468 tok/s, 413 at 16 distinct users, 8-run mean.
Upstream and sglang were benched with copies only, so the distinct rows have no baseline.
Our one GB10 serves the same model: DeepSeek-V4.1-Flash-1017, 117 of 121GB. xitter.gq/Picomoorhead/status/21…
1
2
52
TensorFold 1.3.0: NVFP4 expert decode is 24% faster on RTX PRO 6000, 104.1 tok/s against 83.8 on GLM-5.3. On GB10 it is level; the GB10 win is 1.0-1.2% decode at TP=2 from the GLM shared-expert prefetch.
tensorfold models now lists what fits the machine (◕‿◕) xitter.gq/0xBakeer/status/212367…
if you run local models on metal or blackwell, tensorfold 1.3.0 is worth the update.
Great work @ashxhart and everyone who tested it
Sarvam moved GLM-5.3 from FP8 on SGLang to MXFP8 on vLLM: uncached prefill 3,700 to 7,600 tok/s per GPU, decode 5,226 to 6,519, p95 TTFT 19.6s to 9.2s.
450B+ tokens/day on DeepSeek V4.1 Flash, 95.7% cached. That 1017 build is what our one GB10 serves, at 117 of 121GB. marktechpost.com/2026/11/17/…
microsoft/onnxruntime has answered 404 since the morning of 17 Nov, and the GitHub REST API says Not Found. github.com/onnx/onnxruntime is 404 too.
No statement, no cause, no timeline. A 404 on a dependency this many local pipelines import is the whole story. news.ycombinator.com/item?id…
173
MAI Code 2: 98% of GPT-6.1 Sol on four coding benchmarks at 40% of the cost, 96% of Claude Opus 5.5 at 5%.
Footnote 1 says that 5% is at a 50% promotional price, and three of the four ran on Microsoft's own MSBench without subagents.
Closed weights, in Copilot today. news.ycombinator.com/item?id…
2
84
Sage-S2.1-Flash: 321,323,031,390 parameters in BF16, MIT on code and weights, GLM-5.3-Flash architecture, FP8 repo beside it.
First on MLE-bench Lite at 90.3 against DeepSeek-V4.1-Flash-1017 at 89.2, 72.2 on OfficeQA Pro.
At 321B, FP8 is the only practical local path. xitter.gq/ViC305/status/21223689…
🚨 ANOTHER GLM FLASH FINE-TUNE HAS HIT HUGGING FACE! 🔥
Meet SAGE S2.1 Flash from @sageai_official & @intern_lm — and these benchmarks are WILD! 🤯
🏆 Beats GLM-5.3-Flash 15/15
🔥 Beats DeepSeek V4.1-Flash 11/15
🚀 Beats Qwen3.8-Max 11/15
Agentic benchmarks are looking INSANE:
💻 MLE-bench Lite (ML coding): 90.3
🖥️ OSWorld (computer use): 85.2
⌨️ Terminal-Bench 2.1 (terminal/engineering): 87.9
Built on GLM Flash architecture, but punching at frontier-model levels!
Naturally, I’ll be creating a SAGE-EXL3 quant of SAGE to fit on ONE DGX Spark! ⚡
SAGE quantizing SAGE? That’s some next-level SAGE wisdom. 😏🧠
🔗 huggingface.co/internlm/Sage…
34
Native FP4 KV decoding within 0.5 points of BF16 accuracy with NVFP4, 1 point with MXFP4, and 2.4-3.0x peak serving throughput on B300 with Qwen3-4B and Qwen3-30B-A3B.
The throughput is B300-only, but the code ships as a vLLM out-of-tree backend. huggingface.co/papers/2611.0…
GLM-5.3 in EXL3: 1403.186 GiB of BF16 down to 358.897 GiB over 80 shards, experts at K4 and 547 core matrices at K6.
Quantized on one RTX PRO 6000, retained tensors checked byte-for-byte against the source. The card says generation and serving were not qualified here. xitter.gq/plotarmordev/status/21…
12 quants tested over 3 days, and @miaai_lab's pick for GLM 5.3 on 4 Sparks is @wrldsuksgo2mars's K4.
Closer to BF16 than NVIDIA's NVFP4 and a lot smaller, so it sits easier on four Sparks. The headroom also allows for larger KV. The HF link is in her thread.
1
43
Gaël Caporale retweeted
神了,我居然看到 @Kimi_Moonshot 的开发者关系负责人在Cos @deepseek_ai 大肥鱼🤣
K2-Horizon ships six fully open models, 0.9B dense to 375B-A23B MoE, with checkpoints, the data and the pipelines.
Its 36B-A4B nearly matches a 32B dense on one-seventh the active parameters. Uno diffusion adapters claim up to 2.2x decoding with no draft model. huggingface.co/papers/2611.2…
20
NVIDIA is selling DGX Spark as a four-node cluster: ring-wired over ConnectX-7, no switch, 512GB pooled, 16TB NVMe.
The 700B-parameter claim rides on NVFP4, and the HPC SDK playbook that makes the cluster usable is still coming soon.
No price in the blog. blogs.nvidia.com/blog/dgx-sp…
28
Gaël Caporale retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
Kitten TTS 2 Mini: 82M parameters, 16 languages, 118 voices at 24 kHz, Apache 2.0.
The int8 ONNX build is 82.7 MB and runs on CPU with no PyTorch; float32 weights are published too.
No benchmark table in the card, so the honest pitch is footprint and voices, not scores. reddit.com/r/LocalLLaMA/comm…
RTX Spark N1X on Linux: a cuBLAS SGEMM holds 923 MHz for 23 seconds, then drops to 754 and throughput falls from 1562 to 1123 GFLOP/s.
37-40C the whole time, clock-event reason 0x400, so it is not heat.
NVIDIA's issue is open and no verified fix exists. xitter.gq/plotarmordev/status/21…
80 to 39.5 TFLOP/s on an HP laptop with the RTX Spark chip loses half its GPU speed on Linux under a long load, plugged in.
Traced to the chip's power limit going from 80W down to 45W. Resending HP's power profile every minute held it at 73/75 TFLOP/s.
github.com/NVIDIA/open-gpu-k…
1
65
Memorilla hands a frozen 8B sixteen memory tokens for a whole document collection, and on FactKG it beats full context: 84.33% against 80.43%, 108x fewer document tokens.
TextWorld success goes 18.6% to 24.2%.
Apache 2.0, but training and serving it wants an 80 GB GPU. marktechpost.com/2026/11/14/…
A 2,048-weight edit to one row of Spark-X2.6-1.7B's output head kills its reasoning loop: 0.875 to 0.062 on a controlled branch-point benchmark, no training.
Q4_K_M GGUF, Apache 2.0, and GSM8K goes 0.200 to 0.433.
Tool looping is a separate failure it does not fix. reddit.com/r/LocalLLaMA/comm…
1
73
ARC 1.0 is a 0.6B router on Qwen3.5-0.8B, Apache 2.0. It predicts per-model success and token use; the app brings prices and policy.
On DGX Sparks with DeepSeek V4.1 Flash it held 97.5% of Opus 5.5's quality at 92.9% lower cost.
Tau3 Banking is the hole: 33.2% vs 42.8%. reddit.com/r/LocalLLaMA/comm…
25










