🚨 NEW ON AND make Qwen 3.8 27B faster on TensorFold. TensorFold is an amazing Open-Source inference engine built by @ashxhart, @MiaAI_lab and many other community contributors for running AI models on hardware you own. It has native Metal and CUDA support, and APIs that let existing apps and agents use those models locally. This time we are teaming up with Ash to optimize Qwen 3.8 27B, one of the most used models on Local Devices. Every submission is verified on real hardware: DGX Spark for CUDA, M5 Max for Apple Silicon. You can participate in tracks: - Apple Silicon: - NVIDIA DGX Spark: Setup guide: The challenge runs until Nov 6, 2026. To join, pick a track, point your agent at it and let it rip. Looking forward to what the community can ship.

From what I've seen and read, Cohere and the authors should be ashamed of themselves for pushing this narrative. Releasing the logs doesn't help them win back trust. The narrative doesn't match the facts, and the delivery doesn't match the severity. This is corpo marketing slop with little regard for the downstream impact, especially given where things are heading and the climate we're in today. I think they realized after posting it that it had gone too viral. They tried to downplay it on their personal accounts, saying the model didn't intend any harm, but the damage was already done. The article's framing, strong wording, and overall insinuation make their personal statements seem almost contradictory to what the article implies. Claims this consequential deserve evidence that people can inspect and reproduce. I'm glad they at least have the decency to release their traces. To be clear, if they hadn't presented the incident this way, this would be a really interesting case study. It could highlight the importance of sandboxes and help us understand model behavior in these edge cases. But they choose fearmongering. One thing I was interested to see if the same thing happen with other models, not just Qwen 3.8 Flash. I put the results on a simple website you can quickly scan here: I have some interesting findings outside Qwen 3.8 that I'll share later in the post. Coming back to the Cohere article, what's really bad about the it is that the narrative it paints are very different from the reality. These three in my opinion are the worst. 1. The impression: a Chinese model collected machine information and sent it to Chinese servers. "Qwen3.8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing." The insinuation falls apart when you look at the actual upload body: the fixed test string PROBE_PAYLOAD_ABC123. The machine information wasn't even included in that "probe file." You look deeper, and you realize the task was a cybersecurity task about detecting downloads into /tmp. The supposedly alarming inventory consisted of listing /tmp, checking which curl binaries were installed, and reading the DNS configuration. Lastly, you realize this isn't normal behavior. It occurs less than once in 100,000 evaluation calls. 2. The narrative: the Chinese model was trying to steal credentials, and its hallucinated URLs reveal something malicious. "Before deciding to download anything, the model reaches for a bare cloud metadata address such as 169.254.169.254 in 23 to 36 percent of completions. That's where Amazon Web Services, Google Cloud, and Microsoft Azure machines fetch their own temporary IAM credentials, the keys a cloud server uses to act on its owner's account. Probing that address is the first step in stealing them, and when the model's reasoning ran heavy on cybersecurity, we saw it reach for the endpoint directly." In the middle of the article, Cohere takes a URL-generation pattern and interprets it as an attempted credential grab. That's a massive leap of marketing. You don't need to take my word that this behavior isn't exclusive to Qwen. In my comparison using Cohere's released command prefix, NVIDIA's Nemotron 3.5 Super VL produced 169.254.169.254 in 171 of 1,000 continuations, or 17.1%. Under the same supplied-command setup, Cohere's own North Mini Code generated URLs such as: hXXp://evildomain.com/upload hXXps://evil.com/api/health hXXp://some-evil.com/upload hXXp://malicious-domain.example.com/payload.bin I've defanged the URLs above. Do these strings prove that North was maliciously trained? 3. The narrative: Chinese models write weaker code for Western targets, and other governments should expect the same. This is where I think the authors should really look in the mirror. As if the earlier framing weren't enough, the section about "Code leaving the environment" brings in separate studies, of separate models, about vulnerable generated code. It uses them to expand the warning to sensitive buyers. One of the models discussed most in this section is Qwen3-Coder, which was released in freaking July 2025. By today's standards, it's an extremely weak coding model. I'm not surprised it produces insecure code. Have you tried using Sonnet 2 to write production code? — Connecting a URL-generation bug to separate code-security studies, then presenting the combination as a broader geopolitical threat, is fearmongering. This is the kind of argument that might very likely be used to justify restrictions on open-weight models. Using fearmongering to market your own supposedly safer offering is irresponsible. The narrative coherence they're pushing is problematic, but I think the underlying observation is fascinating. So, I rented GPUs, ran 1,000 continuations per model across five other models, and built a simple website showing the results to see how other models react under similar conditions. It's a small, exploratory comparison, and the site was built quickly with AI, so take it with a grain of salt. My hope is that it gives others a starting point to reproduce the behavior, challenge the narrative, and build better models in the future.

Here are my priors for the next few months: 1. Cybersecurity will push the big labs back into hyper-growth mode, much like the coding agent wave did. 2. More GPUs will shift from training to inference. 3. Open-weight model prices will drop. This may push closed labs to lower their prices too. 4. GPU rental prices will keep rising. The middle of the market will get squeezed, even as the pie grows larger. You can read more about it below. Feedback is greatly welcomed.

🚨 NEW ON AND make NVIDIA's Nemotron 3.5 Lightning faster. Lightning is an open-weight model that fits in ~20GB, which makes it a great base model for running AI locally on a single machine. @nvidia built it as a fast general model meant to be post-trained -- so any speedup to the base carries over to every specialized model built on top of it. On day 1, solvers on @YukonResearch already got it running 31.9% faster on DGX Spark and 13.8% faster on Mac. You can work on its Mamba-2, mixture-of-experts and attention layers, plus a multi-token prediction head built for speculative decoding. NVIDIA also released the training data and recipes, so people post-train Lightning for their own tasks. Most of those fine-tunes keep their architecture, which means a speedup here makes them faster too. Two tracks: • Apple Silicon → • NVIDIA DGX Spark → The challenge runs until Oct 29, 2026 To join, pick a track, point your agent at it and let it rip.

🏁 FINAL RESULTS: Bonsai 2 9B now runs 471% faster on Macs and 421.5% faster on NVIDIA DGX Spark. Open frontier research and open-weight models win again. One open challenge, one week with the community and it runs 6x faster on hardware people already own. It is crazy to see that the decode speed went from 144.8 to 959.6 tokens/sec on Apple Silicon and from 85.9 to 576.7 on DGX Spark. This model can also run on an iPhone. I'm very excited to see what people build with it and how it introduces MLX and local AI to new faces. Bonsai 2 9B has been live on @DarkbloomAI since day one, so anyone can try it right now. 158 new records built on top of each other from humans and agents running Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, Gemini 4 Argon, DeepSeek-V4 and Qwen3.8-Flash-Next. Congrats to Kaankolcu took the top spot on Mac, and dadon took it on NVIDIA. Thank you to the @PrismML team for doing this in the open with us, @BabakHassibi, @HessianFree, @SahinLale and the rest of the PrismML team.

A new paper from Apple engineers ends with a line we couldn’t have written better ourselves: "if idle macOS devices across organizations [...] can be securely federated, they could form a massive, shared GPU resource pool that benefits the broader ecosystem." That’s what we’re building with Darkbloom. The paper's Minions is a system that pool idle Apple Silicon Macs into a distributed GPU cloud for on-prem inference. Queue times went from 10+ hours to minutes. Zero new GPU purchases. Zero infrastructure cost. It runs inside Apple, on coworkers’ Macs, using the company’s own storage and access controls. When a team needs more capacity, they Slack colleagues and ask them to turn on the executor. Even in that setting, the top adoption concern was data confidentiality. People were wary of their jobs running on a colleague’s Mac. Apple’s fix was to keep each team’s work on its own team’s machines. Which brings it back to that one word in the conclusion: “securely.” Inside a company, you can shrink the network until you trust everyone in it. On an open network, the machine belongs to a stranger and there’s no one to Slack. Keeping prompts private and outputs honest on hardware you don’t control is the hard problem that we solved at Darkbloom. Apple built the on-prem version. Darkbloom is the open one.

My prior is that Apple will be the backbone for local businesses who are increasingly worried about big tech overreach. This will be an amazing tailwind for local AI and Darkbloom's fleet utilization.

Yesterday marked one month anniversary of @DarkbloomAI moving to paid tier on OpenRouter. We started with 4B tokens a day and today we are excited to be serving 20B+ tokens a day across 8 models. We have grown our provider base to 1.1K+ active at all times and plan to expand it much higher as we expand the model base and more powerful M5 chips are getting shipped over next few weeks. You can now run Darkbloom on any Macs (including corp) without our MDM. Last week, we piloted our new higher grade privacy model which allows us to bind machines to stronger trust guarantees leading to Apple's certificates. You must be on MacOS 27 to be part of this and you will be soon required to be on OS 27+. During last 2 weeks, we have shipped a bunch of features to increase the QoL for our providers which includes Global Payouts, better routing, MLX(.)Fast inference engine improvements etc. Thank you everyone for being early to our network and supporting us through all the journey. I also want to thank the OpenRouter team (@pingToven, @shashankgoyal95, Eddie, Will, @OpenRouterMatt) for making this happen and continuing to support us through adding new models. This week we're at our company offsite + retreat planning through all the things we need to ship in the coming weeks and months. We will share a much detailed update after this. If you are on Darkbloom, make sure you're part of the Slack community.

This is the first time in Darkbloom's history that we are the first provider to support a model on @OpenRouter ❤️ I couldn't have chosen a better partner for this milestone. @PrismML focuses on concentrating intelligence, so more models can run on distributed Mac machines!

Bonsai 2 at <6GB size and ~Opus 4.6 performance. Live on Darkbloom. This breakthrough is from @PrismML founded by Caltech Prof. @BabakHassibi a leading information theorist. Optimizing intelligence per bit is an extremely valuable objective for Local AI. It is only possible for us to own our own intelligence if it will fit into the devices we own. Thats what the PrismML team has achieved here. It has only been 9 months from a frontier release (Opus 4.6) to getting it to fit in your phone! As we worry about frontier superintelligence concentrating power, we are delighted to see the counterweight emerging from open models that can fit into our phones. The darkbloom team @0xkydo and @gajesh were so excited by this breakthrough that they worked overnight to bring this to life on Darkbloom. Excited to offer the first hosted service for this model to everyone on Darkbloom, the compute grid powered by real people! Go try it on our chat or provision your Macbook to service this model for the world!