I'm building my assistant in Hermes. No, I won't blindly trust it ◆ I teach you the way of responsible AI: your data, your rules ◆ PhD in computer science ◆ Dad
Germany 🇩🇪
Joined January 2022
- Tweets9K
- Following1.3K
- Followers5K
- Likes17.5K
Dr. Daniel Bender retweeted
What a crazy week for open-weight LLMs
DeepSeek-V4-Pro is a 1.4T model so not something I'm going to be running locally but having the choice, and the ability to literally own it, is where open-weight models win.
We’re launching DeepSeek-V4-Pro today! 🚀
🔷 Major Agent upgrades with strong production gains!
🔷 Flexible reasoning effort for V4-Pro & V4-Flash: low for simple tasks, high for daily Agent workflows, max for complex tasks.
🔷 Native OpenAI Responses API support, optimized for Codex with one-click setup.
V4 Pro is now available on app/web. Try it via “Expert Mode”.
V4 Pro is also available via API. Model names remain unchanged—please refer to the API docs for setup details.
Dr. Daniel Bender retweeted
the year of the linux desktop
Linux Desktop Market Share Surpasses 10% in North America
@bimedotcom @Khulood_Almani @theomitsa @Eli_Krumova @sulefati7 @NathaliaLeHen @IanLJones98 @TheAIObserverX
linuxiac.com/linux-desktop-m…
Dr. Daniel Bender retweeted
True to their word, this is now the world's most powerful open-weights model. Not to mention, one of the best Frontier models out there.
There is no back-tracking, dumbing down or banning this model, it is for the people.
Thank you Zhilin Yang and MoonShot.
Releasing the model weights and technical report of Kimi K3.
Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.
New model architecture: 2.5x the intelligence per unit of compute, not just more params.
Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.
Model weights: huggingface.co/moonshotai/Ki…
Tech report: github.com/MoonshotAI/Kimi-K…
Tech blog: kimi.com/blog/kimi-k3
Dr. Daniel Bender retweeted
GPT-5.6 Sol, Terra, and Luna now top @WolfBenchAI. Sol is the strongest model I have ever tested. (Fable scored lower because it refused too often.)
The most interesting part is what happened around the score: Codex used 51% fewer tokens than Hermes while achieving a 2.25 pp higher score.
GPT-5.6 was also cheaper than GPT-5.5 with Codex and T2, but 47% more expensive with Hermes.
The agent changed the economics. 👇
GPT-5.6 Sol took the top three spots on WolfBench:
Codex: 86.74%
Terminus-2: 85.17%
Hermes: 84.49%
So far, so benchmarky. Then I checked the tokens.
With the same Sol model at maximum reasoning, Codex used 83.8 million tokens per run. Hermes used 170.9 million. Twice the tokens for a 2.25 percentage-point lower score.
Hermes recorded $173.41 per run. Codex comes out at an estimated ~$89.79-$95.06. (That Codex number is reconstructed from token usage, not billed spend, but even its upper bound is 45% lower.)
The GPT-5.5 comparison makes it even weirder: With Codex, GPT-5.6 gained 8.09 points, used 20% fewer tokens, and appears roughly half as expensive. With Terminus-2, it gained 8.31 points while cost fell 26.5%. With Hermes, it gained 10.34 points - but used 68% more tokens and cost 47% more.
So is GPT-5.6 cheaper than GPT-5.5? Depends entirely on the agent.
One last oddity: Across 15 Sol runs, the "configure git webserver" task was never solved. Terra solved it once. Luna solved it every time. The smaller variants are not simply weaker copies of Sol.
My takeaway: Codex + Sol is the strongest full-agent stack we tested. Terminus-2 + Sol is the cleanest reference baseline. And for Hermes, I would seriously consider Terra: only 3.6 points behind Sol, but 44% cheaper.
Benchmark the stack, not just the model! And run it more than once.
Full results on WolfBench.ai:
Dr. Daniel Bender retweeted
I'm already a bit tired of the whole US government AI regulation drama. I was considering to write a big post about it, but I don't think I have it in me because everything important has already been said.
Let me just emphasize the absolute importance of open source models. Now more than ever they are essential, if we don't want to risk getting locked out of capable AI models in the future.
GO OPEN SOURCE AI.
Dr. Daniel Bender retweeted
Today, frontier AI online is faster & smarter, so I use mainly that. Long term, local AI only needs to become good enough. Once it just works, I want my own trusted AI running locally, calling online models when needed - while providers focus on global problems, not our inboxes.
Dr. Daniel Bender retweeted
There's FOMO: Fear Of Missing Out.
There's FOMAT: Fear Of Missing Agent Time.
And then there's FOMUT - the next level of agent neurosis: Fear Of Missing Unused Tokens.
> That moment when the limit resets while unused tokens remain - and a precious resource simply evaporates. 💨
ALT FOMUT: Fear Of Missing Unused Tokens
Dr. Daniel Bender retweeted
Today's 🦞 @openclaw tip is about a very useful but little known slash command - and a common problem that's easy to overlook:
/context list
This command shows you what's taking up your context and how large your injected workspace files (AGENTS, SOUL, etc.) are. Here's what many don't realize: by default, each file can have up to 20,000 characters injected, but older versions capped the total for all files combined at 24,000 - newer versions have now increased that limit to 150,000.
Although truncation is less of a problem with this higher limit, it's still smart to check your maximum values and confirm that all workspace files are fully injected (injected chars == raw chars). If not, or if you're approaching any limit, adjust agents.defaults.bootstrapMaxChars (per file) or agents.defaults.bootstrapTotalMaxChars (combined). Remember, this counts characters, not tokens - so 150,000 characters equal roughly 50K tokens.
ALT OpenClaw /context list
Dr. Daniel Bender retweeted
A visual breakdown of the results from my extensive evaluations, benchmarks, and in-depth analyses of all the major LLMs:
ALT Visual breakdown of results from extensive benchmarks and in-depth analyses of the major LLMs
Dr. Daniel Bender retweeted
The harness is the most underexplained piece of agentic AI. Model gets all the attention but the harness — tool call structure, state management, error handling, retry logic — determines production reliability. Most agent failures are harness failures, not model failures.
Dr. Daniel Bender retweeted
Now MiniMax... Open source AI does not sleep - keine Atempause. Minimax 2.5 released. The king is dead, long live the king. Crazy times. I'm very curious in the next one or two weeks to see in which areas these three SOTA open source models shine. Kimi 2.5, MiniMax 2.5, and GLM 5.
Kimi has more experts than GLM, but they have less parameters. AFAIK it should a broader intelligence while GLM should be able to go more deep. About MiniMax, I have no idea about its architecture.
xitter.cf/MiniMax_AI/status/2021…
Introducing M2.5, an open-source frontier model designed for real-world productivity.
- SOTA performance at coding (SWE-Bench Verified 80.2%), search (BrowseComp 76.3%), agentic tool-calling (BFCL 76.8%) & office work.
- Optimized for efficient execution, 37% faster at complex tasks.
- At $1 per hour with 100 tps, infinite scaling of long-horizon agents now economically possible
MiniMax Agent: agent.minimax.io
API: platform.minimax.io
CodingPlan: platform.minimax.io/subscrib…
Dr. Daniel Bender retweeted
I couldn’t agree more. OpenClaw is almost a first ChatGPT release moment. Almost everything it does I’ve been able to do with my own agentic code, prompting, code (vibe-code of course), but it’s just so much faster.
OpenClaw is to Gen-AI what vibe-coding was to code. I’ve installed entire platforms from a train with a prompt on Telegram. I’m literally buzzing with ideas!
Dr. Daniel Bender retweeted
Honestly I wanna release this beast ASAP — I’m dying to go back to my hometown for Spring Festival 😂
But the more training compute we put in, the more it keeps rising.
Painfully happy problem. We hear you guys. M2.5 soon.
Dr. Daniel Bender retweeted
There was a recent change from "call-based" to "token-based" usage for Kimi.
Nevertheless, it is not looking promising. I already used nearly half of my weekly tokens of the coding plan in 2.5 days. 🙈
Besides that, I really like the responses generated with Kimi K2.5. 💪
Dr. Daniel Bender retweeted
People keep asking: "Yeah, but what can I actually do with ClawdBot?!"
One thing I've been recommending: when you get an idea for a project, just dictate it to Clawd aka OpenClaw, ramble about all the cool ideas you have. Tell it to put it in your list of ideas.
Then tell Clawd that every night while you're asleep, it should choose one idea and it should build you a little demo that you can wake up to.
Dr. Daniel Bender retweeted
#1 open-source, #6 overall. Close to Gemini 3 Pro, beating GPT-5.2.
Solid execution, strong design sense, reliable vibe.
For real web and app builds, you can always trust MiniMax-M2.1.
🚨Code Arena Update
Minimax-M2.1 debuts at #1 open model on WebDev leaderboard and lands #6 overall with a 1445 score, tying with GLM-4.7.
These scores come from Code Arena, where models build websites, apps, and games from a single prompt.
Congrats @minimax_ai on this latest milestone!
Dr. Daniel Bender retweeted
all my agents are running locally right now
> yup, 100% backed by my GPUs
> claude and chatgpt both have been nerfed today
> just a preview of what happens when someone else is controlling your ai
yet another reason to Buy a GPU & run your AI locally
Dr. Daniel Bender retweeted
Love it and would use it as my main AI interface. Only gripe: unlike Gemini CLI, it's not open source. That's a big miss - it's built on open-source VSCode, and I'd happily contribute.
So much progress could happen fast. So my plea:
Please make @antigravity open source too! 🙏
Introducing Antigravity (@antigravity)
An agentic development platform, including task lists, verifications, browser integration, and more.
Try it! antigravity.google/
Blog antigravity.google/blog/intr…
