Goals: Power to individuals. Improve my surroundings. Means: Blockchain. Bitcoin. Strong typing. Community.
Vienna
Joined October 2009
- Tweets783
- Following698
- Followers687
- Likes865
Just realised @pidotdev and @DeepSeekHarness are closer than expected. DSH is the Android to PI's Linux
Chonky model, weird physics in the video. Exited for the weights release, But making this run fast will be a challenge with 49B active params. Dual M5U 512 around 40-50 tokens/sec max ..
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
Andreas Petersson retweeted
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
🤖 Made with AI
Andreas Petersson retweeted
You paid a Lightning invoice, but the receiver says the sats never arrived? 🤔
Your LN wallet has the proof: the preimage.
But unlike on-chain payments, there’s no transaction to point to in a block explorer.
How do you explain a 64-character hex string to a non-expert? 🧵👇
ok @thsottiaux the Tuesday has arrived even in Attu Station. my 100€ codex credits from Astra Commons vanished in less than 15 minutes. @codex_resets let's go, we got important work to do.
Trying @earendilworks Radius: every built-in tool bills a flat $0.05/call. web search, fetch, image gen and file describe. Parallel search lists $0.001–0.005 and describe-file should be near-free, that feels steep. Maybe I'm missing something? Happy to be corrected.
Thanks @antirez for the recent ds4 updates and YouTube video. Phenomenal work! I retrofitted experimental DeepSeek V4.1 steering to the CLI/server and published three adapters with provenance. Behavioral validation still pending.
github.com/apetersson/ds4/tr…
huggingface.co/apetersson/De…
WOW!! DeepSeek V4.1 Flash weights are here! 🔥
Learned n-gram memory is a smart way to expand model capacity without scaling dense compute. NVMe offload could make that extra capacity practical for local inference.
huggingface.co/deepseek-ai/D…
Qwen3.8-Flash-Next: iMatrix Q5, the sweet spot for 128GB systems. Clever n-grams stay BF16 on SSD, freeing RAM for context.
M1 Ultra @8K: 176 tok/s prefill, 32 decode. 1M context tested.
Try it and report your numbers:
huggingface.co/apetersson/Qw…
Your AI workload can burn twice the credits at the wrong hour.
Time zones and daylight saving make cheap windows easy to miss. Offpeak Clock now tracks Tencent and Z.ai GLM-5.3-Flash:
offpeakclock.com/#provider=t…
offpeakclock.com/#provider=z…
DeepSeek V4 Flash Vision Exp, surgically abliterated now handles text+image in DS4. Fixed vision layout/attention, added direct HF loading, and stream MoE from SSD so the 165GB model runs on 128GB Macs. A 128GB-friendly quant is coming soon. huggingface.co/apetersson/De….
What is model abliteration, I was recently asked. Won't it negatively impact the model? Think about a brainstorming session with 3 enigneers 1 designer and 1 laywer. Abliteration removes the lawyer.
Okay, what exactly is GPT “Daybreak Blue”?
Gave it a fairly niche smart-contract/security task and it casually produced one of the most complete implementation plans I’ve seen.
Then I started getting “model at capacity” errors.
Anyone else testing Daybreak yet? 👀
This is what competition looks like. I love how almost every node on the pareto frontier is a different icon. Time to move it up to the top right corner! @GeminiApp @OpenAI
DeepSeek-V4-Flash (High) by @deepseek_ai has reshaped the cost-performance Pareto frontier in Agent Arena, with a $0.024 median cost per task!
It lands to the right of GPT-5.6 Luna (xHigh) which has a $0.026 median cost per task, and to the left of DeepSeek-V4-Pro (Thinking) at $0.020. It's the model that delivers a positive net improvement at the lowest price point on the chart.
Price per task is based on real-world Agent Mode usage. We calculate each task’s cost from tokens consumed and the model’s pricing, including cache hits and misses.
