While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: https://t.co/TZSugXi2ZE
Hanging out at the @cognition hack working with #devin
4hrs to whip up a sales coaching app, anyone not optimizing internal org structure with AI is missing out
Thx @launchloop_xyz for organizing this!
GPT Astra just cooked Claude Fable 5.1 — for a lower price!
OpenAI just rolled out GPT‑6 Astra — we put it against Anthropic's Fable 5.1
Both models got the same four briefs: Earth colliding with Mars, a tall ship sucked into a giant whirlpool, Godzilla leveling downtown, a mega-tsunami swallowing a city. One self-contained HTML file per scene, one shot, no edits.
GPT Astra: 25,159 tokens, $1.67
Claude Fable 5.1: 37,767 tokens, $2.51
both models live on aimlapi[.]com — 1000+ models, one key.
Muse Glimmer one-shot 5 games!
@AIatMeta dropped Muse Glimmer. So we ran it against Muse Spark 1.2, on five one-file game demos, each one written to play itself: Tetris, a top-down pixel street race, a rooftop web-slinger, a blue hedgehog platformer, and Flappy Bird.
the setup:
• Muse Glimmer 30B — via aimlapi[.]com. cost: $0.02
• Muse Spark 1.2 — via aimlapi[.]com. cost: $0.11
what we saw:
• every Glimmer scene runs and plays itself. it's a reasoning model, so it burns more tokens — but the output is clean. its Tetris even scores each placement to keep lines clearing, and the blocks burst into pixels on every clear.
• Glimmer beat Spark at Flappy Bird — its bird flew through the pipes cleaner.
• but it's not perfect: in the racing game its cars sometimes slip off the road, and its graphics are simpler and blockier than Spark's.
glad to see Meta back in open source — and we're waiting for Spark 1.2 weights next 🙌
Seedance 2.5 just cooked Veo 3.1 – for half the price!
to celebrate, we're giving away $10 in api credits to 20 of you to try seedance 2.5 – just rt/qrt this post, winners will be picked randomly on August 10
@BytePlusGlobal 's seedance 2.5 just dropped. we ran it with google's Veo 3.1 and honestly, it wasn't much of a fight
the test: 5 prompts, one shot each, 4s, 720p, native audio, same exact prompt to both models: a tornado grinding through downtown, a tsunami swallowing a megacity, slow-mo liquid physics, a 180° camera orbit, and a neon sign with text
cost for the 5 clips:
Seedance 2.5 – $6.08 ($1.21 per clip)
Veo 3.1 – $10.40 ($2.08 per clip)
observations:
• seedance is 42% cheaper and it's the better one.
• seedance keeps geometry locked through the orbit shot – buildings, the cellist, the station all stay solid.
• disaster scenes are the closest call: veo throws more debris, seedance stages the shot better.
same prompts, almost half the price, better result. bytedance shipped a monster – and it's live on aimlapi[.]com!
Atomic Chat signed the Open Weights letter!
We believe everyone should be able to run AI on their own device. When a model is open, thousands of teams fine-tune it, quantize it and build new tools on top of it. Together we move AI faster than any one company ever could
We've tested the Qwen 3.8 Max preview and it absolutely rocks!
Each scene is built in one shot, a single Three.js file with nothing else added:
• Odysseus' war galley riding the waves
• Trojan horse
• Greek war helmet
Qwen managed to execute the task faster than the other models, while output quality is relatively the same across all of them. If the pricing stays close to Alibaba's previous flagships, we'll get one of the best speed/quality ratios on the market — perfect for running agents at scale.
Soon live on AI/ML API.
First local agent to beat Hermes on benchmarks!
✦ runs Qwen, Gemma, Llama via llama.cpp
✦ stable-prefix caching keeps sessions cheap
✦ TurboQuant cuts the KV-cache 6.4× smaller
✦ 37 tasks solved vs Hermes' 31 on GAIA Level 1
Open source on macOS, Windows & Linux 👇
New Kimi K3 beat Opus 4.8 at making retro games for 2x cheaper!
We gave three models the same task: make three old arcade games that play by themselves. Each game is one HTML file with a bot that plays it
Prompts:
– Road Fighter
– Battle City
– Q*bert
Outputs:
Kimi K3: 18.4K tokens, $0.28
GPT-5.6: 18.1K tokens, $0.28
Opus 4.8: 21.3K tokens, $0.54
Kimi K3 made the best games. Its Q*bert is the best of all three. The player jumps on the cubes, paints them and runs from the purple snake. Its Battle City looks just like the real game and plays right. The tanks move, aim, shoot and break the walls. GPT-5.6 could not handle the cars. Its Battle City broke too. The tank died on its own and the base got hit. But GPT had the nicest textures of the three
GPT 5.6 Sol & Terra just made Fable 5's pricing look like a joke!
We gave OpenAI's top models (Sol, Terra) and Anthropic's top models (Fable 5, Opus 4.8) the same 3 prompts:
• Supernova boom
• Meteor hitting a city
• Solar system model
The bill:
Sol: $4.77
Terra: $1.24
Fable 5: $9.94
Opus 4.8: $2.46
Outputs came out surprisingly close. The prices didn't. Fable 5 cost unreasonably more than everything else, with Sol not far behind. Considering latest Fable 5 nerf, it's hard to see what you're paying for.
Open-weight LongCat 2.0 matched GPT-5.5 level on agentic game dev for $0!
We ran Meituan's LongCat 2.0 against cloud frontier GPT-5.5 in @kilocode CLI with their agent. Same task for both - build a retro Duck Hunt game in one game.html, improved over 3 agent iterations with duck waves, ammo and physics
Outputs:
LongCat 2.0: 70.3K tokens, $0.00
GPT-5.5: 64.9K tokens, $0.65
LongCat kept up on graphics, physics and game logic. Ducks fly and fall when hit, the dog fetches them, ammo counts down, the waves keep coming. Both ran clean and nothing clipped. The only difference was the bill - GPT cost $0.65, LongCat ran local for $0
Fable 5 totally crushed our new contest, but it cost 6x more than Opus 4.8!
We gave 4 models the same prompt: build three self-contained HTML5 canvas scenes with real physics demos
Prompts:
— A train derailing off a broken bridge into the water
— Two cars jumping off ramps and colliding mid-air over a canyon
— A monster truck crushing a row of parked cars
Outputs:
Fable 5: 62,158 tokens, $3.12
GPT 5.5: 37,753 tokens, $1.14
Opus 4.8: 22,280 tokens, $0.56
GLM 5.2: 36,246 tokens, $0.08
Fable 5 did all three scenes at A+. The crashes looked real, things fell and broke the right way, and nothing went through the ground or floated. GPT 5.5 was the closest to Fable. In the Bigfoot show, we think GPT was even a little better. GLM 5.2 did not win any scene, but it was the cheapest by far. Fable is the best pick for quality, but you pay more for it.
New Claude Sonnet 5 performs at GPT 5.5 level 6x cheaper!
We gave 4 models the same prompt: build three self-contained HTML5 canvas scenes with real physics crash demos
Prompts:
- A car crashes into a brick wall
- A wrecking ball destroys a house
- A catapult throws a rock at a castle wall
Outputs:
Sonnet 5: 15,047 tokens, $0.15
Opus 4.8: 23,063 tokens, $0.58
Sonnet 4.6: 25,824 tokens, $0.39
GPT 5.5: 31,152 tokens, $0.94
Sonnet 5 did as well as Opus 4.8 and GPT 5.5 on all three tests. In the wrecking ball test, it beat Opus 4.8. The cable moves smoothly and every hit connects. In the catapult test, it beat GPT 5.5. The rock always lands inside the wall. Sonnet 5 still needs better detail and graphics. But it used fewer tokens than every other model
If you've been wanting to run AI privately, offline, and for free but terminal/code was the blocker this is for you.
Stop using ChatGPT for everything. Run the SAME quality models on your own laptop. for free, forever, with zero terminal commands.
I just recorded a full walkthrough showing how anyone (yes, even if you've never touched a command line) can download and run the latest open weight LLMs from Hugging Face using llama.cpp wrapped in an intuitive GUI.
In the video I:
- Search & download Gemma 4 12B straight from Hugging Face inside the app (For my NVIDIA RTX 4060 8 GB VRAM)
- Chat with it locally - no API key, no internet required, no data leaving your machine
- Connect it to Hermes Agent in 4 clicks and have it running real agentic tasks on a 100% local model
The tool I used is Atomic Chat. open source (Apache 2.0), free, and built on llama.cpp under the hood:
- Native apps for mac, windows, linux, iOS and android, not just desktop
- One click agent integrations: Hermes Agent, Kilocode, Claude Code, Codex CLI, OpenCode, Cline CLI, Droid etc.
- Full MCP support. plug in tools and let your local model actually do things
- Auto extends context length when you're about to run out. no more silent truncation
- Hardware aware model recommendations, so it tells YOU which model your machine can actually run instead of you guessing and crashing your RAM
- OpenAI compatible local API server if you want to wire it into your own scripts later
This is the local LLM setup you need when you want started ASAP with a great UI: HuggingFace + llama.cpp power, GUI simplicity, and agent ready out of the box.
Download from the link in the comments and let me know if you need any help.
What's stopping you from running AI locally right now RAM, know how, or just never knew this existed?
Email for AI agents integrated in Dify.
Drop @atomic_mail into any @dify_ai workflow. Agent registers its own address in one click – then triage, send, route and summarize
Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding.
Ornith-1.0 spans the full parameter sizes including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. It achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks including:
✅Terminal-Bench 2.1(77.5)
✅SWE-Bench(82.4 on verified, 62.2 on pro, 78.9 on Multilingual)
✅NL2Repo(48.2)
✅SWE Atlas(41.2 on QnA, 42.6 RF, 39.1 TW)
✅ClawEval(77.1)
Post-trained on top of gemma4 and qwen3.5, Ornith-1.0 employs a novel self-improving training strategy in which reinforcement learning is used to generate not only solution rollouts, but also the task-specific scaffolds that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model generate higher-quality solutions in agentic coding.😎
All models are released under the MIT license, enabling full commercial and research use.
📖Tech Blog: https://t.co/qT9N2HYWFn
🤗Huggingface: https://t.co/PRrwqjeBtM
Exa AI Search now in Atomic Chat!
Your local models can finally search the live web and pull fresh info — powered by @exaailabs, trusted by 400,000+ developers, free and open-source