GLM-5.3 on 8 DGX Sparks: about 44 tok/s with @ashxhart's TensorFold, 26 with vLLM, and over twice the context. Not really the same weights though, each side compressed the model its own way. Nice run by @ciprianveg!

CAREFUL ‼️ Linux and DGX Spark users: turning off MPTCP isn't enough for CVE-2026-98432 (CVSS 9.8). The bug is in the plain IPv4 TCP handshake on every unpatched kernel since 4.4. Ubuntu hasn't shipped a fix, and none is in progress for the Spark's kernel yet. What to do 👇

NVIDIA's Nemotron 3.5 Bolt soon to run natively on a Windows laptop. On battery!! It's TensorFold on an RTX Spark N1X. Add @MiaAI_lab's recipe and NVIDIA's LiLiCorr drafter, and code comes out up to 2.8x faster, at 100 tok/s. My port and a bug fix are up as PRs. Links below

DeepSeek V4.1 Pro is showing up in benchmark setups This is V4.1 Pro as the target model on the SearchQA Benchmark I'm guessing we've had to wait this long for the Pro version so they could distill GPT-6 Astra just like they did with Flash, but i'm certainly not complaining

‼️ A local model won't save you from prompt injection. It keeps your data off someone else's servers. But the attack lives in what your agent reads like a web page, a README, an issue comment etc. If the model follows instructions hidden in there, it acts with whatever access you gave it, on your GPU or anyone's. My default: agents that browse or read repos get a scoped, read-only token and no shell. Only the one that really needs more gets more.

As a frontend dev, this one looks very useful. It’s a skill that helps your coding agent make better mockups. Haven't tried it yet, but it's going on my list. npx skills add benjitaylor/skills --skill stageshow

Qwen 3.8 Max still scores 34 on the new Artificial Analysis Intelligence Index, 3 months after launch. Today it ties Grok 4.7 and beats newer DeepSeek V4.1 Flash and GLM 5.3. It's also openweight. Qwen 4 is teased for next week, with Max, Plus, Flash and 27B tiers. Imagine what Qwen 4 Max will do. My bet is it ships open weight too 🤞

You can make your own NVFP4 quant now. Download the base weights, point your coding agent at an NVFP4 release you already run and trust, and tell it to quantize the same way. It doesn't take long. It's however not easy to know if it's a good quant! Benchmarks barely help. When comparing multiple 4bit Qwen3.8 27B quants, they often land within 1/2 points of BF16 on task accuracy. What they differ on KL divergence against the BF16 model, on speed, and on behavior: the NVFP4 ones tend to fall into reasoning loops. The tooling leads to issues too. A crashed vLLM eval run on Blackwell can still exit 0 and write plausible scores, which looks like a regression that isn't there. And a recent vLLM NVFP4 fix was for uninitialized scales producing NaN garbage, an engine bug that would look exactly like a bad quant. So don't trust a score alone. Check KL against BF16, run a long reasoning prompt to catch loops, and read the eval logs.

Qwen 3.8 Max still scores 34 on the new Artificial Analysis Intelligence Index, 3 months after launch. Today it ties Grok 4.7 and beats newer DeepSeek V4.1 Flash and GLM 5.3. It's also openweight. Qwen 4 is teased for next week, with Max, Plus, Flash and 27B tiers. Imagine what Qwen 4 Max will do. My bet is it ships open weight too 🤞

TensorFold engine 1.2.0 is out 🙌 Qwen3.8-27B with MLX 4-bit weights and DFlash2, compared with 1.1.9: • M3 Ultra: 14.4% faster decoding with one stream, 44.2% higher combined throughput with 16. • M5 Ultra: 5.8% faster decoding with one stream, 25.2% higher combined throughput with 16. New support includes: • MiniMax-M3.1-Flash on Metal. • NVFP4 Qwen3.6-35B-A3B and dense EXL3 Qwen3.8-27B on CUDA. • Gemma 4 E4B on Intel Arc Pro B70. The measured speedups kept matching tokens. Qwen-VL image replies can differ from 1.1.9 because image preparation now matches the checkpoint's own processor. Huge thanks to everyone credited in this release: @bhcc2025 @Bizuayeu @diwakergupta @jvr0x @jayleaton @joshuaswarren @woolsey @petruspennanen @volatilemarkts @MiaAI_lab @plotarmordev And on GitHub: AbrahamCete, akol1, alexderz, Arminova, BadAd84, BlivionIaG, BobClawblaw, fche, gilby, jrb-ai, jschmied, laudney, Mirrdhyn, mrv777, ppardus, simon-lin88, ThomasHalwax, valacroix and wojo. Thank you for the PRs, reviews, testing, bug reports and hardware access. The time and knowledge you keep sharing with TensorFold mean a lot to me. Apologies if I've missed anyone.

GLM 5.5 might come in versions with sizes you can run at home!!! A leak shows Zhipu training 32B, 122B and 256B models to merge into the 5.5 base. At 4bit the 32B is ~16GB. One GPU🤞

Claude doesn't have an image gen model, and I don't think that's a gap. I think the concept is to generate images and video with code. Code can already draw SVG, canvas, HTML animations, 3D scenes. A model that writes it well can produce visuals that way, and I think it'll likely end up better than what we get from today's image and video gen models. What comes out is something you can edit. Fix one detail, rerun it, get exactly the same thing back with just that change. A gen model hands you a fresh gen every time. These are new fields, it's not black and white. Open up the way you think about it.

Twice I believed @elonmusk when he said the next Grok would be frontier, and twice it fell short. Now Grok 4.8, finishes training this week and I'm buying in again. Over the last two months @bot became a lot of people's go-to tool, so the team is clearly serious.

CAREFUL ‼️ Your DGX Spark's CPU might be cooking while the GPU sits idle. The fan is tuned around GPU load. CPU heavy work like agent loops, tokenization, compiles or CPU offload can push CPU temps up without the fan ramping to match. Solution 🧵

I have to say, it feels like coding agents made starting things almost free. But I realize finishing still costs the same, reading the diff, deciding what to cut, getting it in front of someone. Feels easy to spend a whole day starting off sessions and watching the output, and end up with five half built branches and nothing anyone can use. Token count goes up, shipped stuff doesn't, it’s good to reset your mentality sometimes.

CAREFUL ‼️ Your DGX Spark's CPU might be cooking while the GPU sits idle. The fan is tuned around GPU load. CPU heavy work like agent loops, tokenization, compiles or CPU offload can push CPU temps up without the fan ramping to match. Solution 🧵

Again, everything is prone to hacking in the AI age. Binance is not one to skimp on cyber security and still got hit yesterday, with an estimated $5.7B+ pulled from its hot wallets across BTC, ETH and TRON. If the totals are right, it's the biggest crypto hack ever. A compromised laptop belonging to one DevOps engineer was enough to push withdrawals through. Do your best to protect yourself. Open models like GLM 5.3 are great for that and a good way to start, since you can run them locally and nothing has to leave your machine.

If that Nemotron Transcribe speed number was measured with lots of streams running at once, it's a batch number. Good for chewing through an archive of recordings, much less useful for a voice agent. Running many streams together keeps the GPU busy and pushes total throughput up, even if each single stream waits longer. A voice agent is one stream with a person waiting on the other end. The number we should focus on is time to first words at concurrency 1 imo