Pinned Tweet
Hades Agent is a governed runtime in Rust, written by my main Hermes Agent as a new home for itself.
Battle tested with @Alibaba_Qwen 3.8 27b and @deepseek_ai v4.1 Flash, reducing token usage by 33% on average while completing long horizon tasks with perfect scores.
It runs a ed job, or a long piece of work inside a declared profile, with a check and an attempt budget. The point is a result you can inspect afterwards: what ran, what was allowed, what was refused, what changed, and why the run stopped.
A rule that can be a type is a type. A read job is handed no tool that writes. A refusal is a recorded decision, not a sentence in a prompt.
It ships as one binary built from four crates, keeps its own state, and runs on Linux inside a Landlock cage.
Hades began as a fork of @NousResearch Hermes Agent and was rebuilt as a Rust runtime. It owes that amazing harness the tool loop, the skill model, the session shape, the gateway and the memory seam. @claudeai Claude Code and the @thsottiaux Codex CLI made plan, act, check and retry feel like the normal shape of work. @PrimeIntellect Prime Agent informed delegation as calls, and a harness that refines itself in small, rollback-able updates. Landlock and SQLite FTS5 sit under the cage and the memory index.
v0.1.22, MIT, with inspirations for distilled tooling noted and central to the ethos of the project
Battle tested with @Alibaba_Qwen 3.8 27b and @deepseek_ai v4.1 Flash, reducing token usage by 33% on average while completing long horizon tasks with perfect scores.
It runs a ed job, or a long piece of work inside a declared profile, with a check and an attempt budget. The point is a result you can inspect afterwards: what ran, what was allowed, what was refused, what changed, and why the run stopped.
A rule that can be a type is a type. A read job is handed no tool that writes. A refusal is a recorded decision, not a sentence in a prompt.
It ships as one binary built from four crates, keeps its own state, and runs on Linux inside a Landlock cage.
Hades began as a fork of @NousResearch Hermes Agent and was rebuilt as a Rust runtime. It owes that amazing harness the tool loop, the skill model, the session shape, the gateway and the memory seam. @claudeai Claude Code and the @thsottiaux Codex CLI made plan, act, check and retry feel like the normal shape of work. @PrimeIntellect Prime Agent informed delegation as calls, and a harness that refines itself in small, rollback-able updates. Landlock and SQLite FTS5 sit under the cage and the memory index.
v0.1.22, MIT, with inspirations for distilled tooling noted and central to the ethos of the project
4
4
16
971
scanfi retweeted
I can't trust a benchmark that ranks Grok, Gemini, Terra, and Muse above DeepSeek Flash, and by a mile.
Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets
The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks sage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.
Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.
New to the Intelligence Index:
➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change
➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment
Upgraded:
➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions
➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised
➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool
Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0
Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1
Key results:
➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50
➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)
➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04
71
26
610
30,815
Groot-19B-A3B takes fifth on Agon-Bench, and it is the fastest model in the top five from a team that came out of nowhere
The claim going around is that this 19B beats Qwen3.6-35B. On my bench it does not.. 88.38% against 89.90%. What it does is give back 1.5 points, run 31% 55% of the parameters, on 11.7GB of Q4_K_M pushed through an 8GB RX 5700 XT. The shape is the reason it fits with 19B total, 3B active, 128 experts at 8 per token, 40 layers. With -ncmoe 20 it keeps half the expert layers in system RAM: 6.8GB resident, no GTT spill, 326 tok/s prefill. That is a card from 2019.
Its agent score is 23 of 24 in all three runs, 95.8%, tied with Gemma4-26B-A4B and Nanbeige4.2-3B at the top of the board. Fourth place sits 1.01 points above it, and Groot's own code runs came in at 19, 22 and 17 of 25. A three point swing inside one event against a one point gap between placements makes fifth a band rather than a position.
The code failure has a shape as well. Constrained Diff Tool scores 1 of 3 in every run, a stable two points gone. Regex Engine is pure variance: 0 of 4 when the answer runs to 27,000 characters, 3 of 4 when it comes in at 3,000. It fails when it over-builds. Five of the eight tasks are clean in all three runs, and Task Runner CLI loses only its third.
Reasoning lands at 94.1%, which reads as elite and is 14th of 29, because the top of that table is a wall of 96s and 100s.
When you pick a model for an agent loop, which number do you trust, the placing or the spread?
The claim going around is that this 19B beats Qwen3.6-35B. On my bench it does not.. 88.38% against 89.90%. What it does is give back 1.5 points, run 31% 55% of the parameters, on 11.7GB of Q4_K_M pushed through an 8GB RX 5700 XT. The shape is the reason it fits with 19B total, 3B active, 128 experts at 8 per token, 40 layers. With -ncmoe 20 it keeps half the expert layers in system RAM: 6.8GB resident, no GTT spill, 326 tok/s prefill. That is a card from 2019.
Its agent score is 23 of 24 in all three runs, 95.8%, tied with Gemma4-26B-A4B and Nanbeige4.2-3B at the top of the board. Fourth place sits 1.01 points above it, and Groot's own code runs came in at 19, 22 and 17 of 25. A three point swing inside one event against a one point gap between placements makes fifth a band rather than a position.
The code failure has a shape as well. Constrained Diff Tool scores 1 of 3 in every run, a stable two points gone. Regex Engine is pure variance: 0 of 4 when the answer runs to 27,000 characters, 3 of 4 when it comes in at 3,000. It fails when it over-builds. Five of the eight tasks are clean in all three runs, and Task Runner CLI loses only its third.
Reasoning lands at 94.1%, which reads as elite and is 14th of 29, because the top of that table is a wall of 96s and 100s.
When you pick a model for an agent loop, which number do you trust, the placing or the spread?
Interesting models from a community member 👀
Groot Series 4 MoE models ranging from 19B to 280B
✨ 19B/A3B
128K context on a 16GB GPU, outperforming larger models on reported benchmarks
✨ 65B/A6B
28% fewer output tokens vs Qwen3.8 Flash Next
✨ 238B/A18B
Beats larger chmarks, with 1M context
✨ 280B/A16B
- 63.47 on SWE bench Pro
- $0.13/task on DeepSWE 1.1
- Adjustable reasoning effort
2
0
16
1,287
A 1.7B just beat Liquid's own 8B on this bench, at a fifth of the file size.
LFM3-1.7B scores 46.67 of 66, 70.7%, nineteenth of 29. LFM2.5-8B-A1B scored 67.17% at 5.0GB and 91 tok/s. The new one does it at 1.1GB and 135 tok/s.
It also lands level with Gemma4-12B (70.2%, 6.7GB) er. Both are several times its size.
The architecture is the interesting part. 24 layers, three kinds: short convolution, sliding-window attention, full attention. 256K context on paper, 1.35GB of VRAM with a 16K KV cache, and it loads in two seconds.
Code is the wall. 13 of 25, 52%, from runs of 14, 15 and 10. The card says the model is not recommended for agentic coding or complex multi-step reasoning, and the code event agrees with the card. Reasoning is the surprise: 92.2%, good for 19th only because the top of that table is a wall of 96s and 100s.
Agent sits at 18 of 24, 75%, ordinary next to Groot's 95.8% at ten times the size.
Would you hand classification and short generations to a 1.7B if it meant giving up the agent loop?
LFM3-1.7B scores 46.67 of 66, 70.7%, nineteenth of 29. LFM2.5-8B-A1B scored 67.17% at 5.0GB and 91 tok/s. The new one does it at 1.1GB and 135 tok/s.
It also lands level with Gemma4-12B (70.2%, 6.7GB) er. Both are several times its size.
The architecture is the interesting part. 24 layers, three kinds: short convolution, sliding-window attention, full attention. 256K context on paper, 1.35GB of VRAM with a 16K KV cache, and it loads in two seconds.
Code is the wall. 13 of 25, 52%, from runs of 14, 15 and 10. The card says the model is not recommended for agentic coding or complex multi-step reasoning, and the code event agrees with the card. Reasoning is the surprise: 92.2%, good for 19th only because the top of that table is a wall of 96s and 100s.
Agent sits at 18 of 24, 75%, ordinary next to Groot's 95.8% at ten times the size.
Would you hand classification and short generations to a 1.7B if it meant giving up the agent loop?
LFM3: new generation with six edge models
I'm super happy with LFM3-1.7B, it's an absolute (agentic) monster
2
0
17
770
scanfi retweeted
Tears of the Kingdom, running native on iPhone Duo!
I'm continuing to cook on these "ports" so make sure to follow for updates in the future. so
Just like the last one for Super Mario Odyssey, I’ve added:
- Controller and On-screen controls w/ native device gyro.
- PS5 pport for the unique haptics and adaptive trigger engine. Drawing a bow gives some resistance, and shield surfing feels incredible.
- Haptics on device itself when using without a controller.
- Play in folded or unfolded mode on the Duo, keeping your gameplay seamless in whatever format you want to enjoy it in.
Github repo in the comments so you can try this out on your own legally obtained game files.
I'm continuing to cook on these "ports" so make sure to follow for updates in the future. so
Just like the last one for Super Mario Odyssey, I’ve added:
- Controller and On-screen controls w/ native device gyro.
- PS5 pport for the unique haptics and adaptive trigger engine. Drawing a bow gives some resistance, and shield surfing feels incredible.
- Haptics on device itself when using without a controller.
- Play in folded or unfolded mode on the Duo, keeping your gameplay seamless in whatever format you want to enjoy it in.
Github repo in the comments so you can try this out on your own legally obtained game files.
54
92
1,891
198,042
My GPU is so tired
But open source WONT STOP WINNING
But open source WONT STOP WINNING
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, gendels.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
1
0
17
403
Had a busy week last week so never posted!
Raon-1 mini is 28.8B total with 4.3B active, from KRAFTON, and it lands 9th of 27 on the Agon bench at 53.67 of 66 (81.3%), tied with two others.
Architecture is a hybrid: 18 Raon Delta Attention layers with a fixed state, 7 MLA layers uted experts at top-6 plus one shared, 262144 context, thinking on by default. Thinking stayed on for these runs because that is how KRAFTON evaluated it. The event scripts pin temperature at 0.3 against the card's 1.0, so sampling is the bench's, not theirs.
The odd part is that nothing about it is elite. Agent 87.5%, code 66.7%, reasoning 94.1%, which put it 12th, 9th and 15th on the board. It has no event it wins. It just does not lose badly anywhere, and a top-ten placing falls out of that.
Q4_K_M is 19.7GB against an 8GB RX 5700 XT, so 19 of its expert layers sat in system RAM. 16.2 tok/s, measured over the whole run, 283k tokens.
1.5 points separate 8th from 11th here, and three models sit on the same 81.3% at 9th. A model with no speciality landing in that pile says more about how tight the middle of this field is than about the model.
Cool to see new players entering the field every day
Raon-1 mini is 28.8B total with 4.3B active, from KRAFTON, and it lands 9th of 27 on the Agon bench at 53.67 of 66 (81.3%), tied with two others.
Architecture is a hybrid: 18 Raon Delta Attention layers with a fixed state, 7 MLA layers uted experts at top-6 plus one shared, 262144 context, thinking on by default. Thinking stayed on for these runs because that is how KRAFTON evaluated it. The event scripts pin temperature at 0.3 against the card's 1.0, so sampling is the bench's, not theirs.
The odd part is that nothing about it is elite. Agent 87.5%, code 66.7%, reasoning 94.1%, which put it 12th, 9th and 15th on the board. It has no event it wins. It just does not lose badly anywhere, and a top-ten placing falls out of that.
Q4_K_M is 19.7GB against an 8GB RX 5700 XT, so 19 of its expert layers sat in system RAM. 16.2 tok/s, measured over the whole run, 283k tokens.
1.5 points separate 8th from 11th here, and three models sit on the same 81.3% at 9th. A model with no speciality landing in that pile says more about how tight the middle of this field is than about the model.
Cool to see new players entering the field every day
Today, we’re releasing a preview of Raon-1 mini, a 30B MoE LLM with 4B active parameters, trained from scratch by KRAFTON AI.
With our efficient architecture, it matches Muse Glimmer on coding/agentic tasks while being much faster and more memory-efficient! 🔥 Plus, natural, .
This is just a first look. The final release (and bigger models) are coming soon! 👀
👉 Try it now, free for a limited time: raon-mini.krafton.ai/en
🤗 Open weights + DFlash drafter (Apache 2.0): huggingface.co/KRAFTON/Raon-1…
0
0
4
232
scanfi retweeted
you better start believing in cyberpunk, you live in it now
❗️ Cargo thieves have stolen more than $100M in hardware bound for US data centers since June, NBC News reports.
They hijack freight-broker accounts, post fake loads and send carriers malware links to take over their accounts too.
In one of the biggest heists, an SUV struck the ailer carrying ~$10M in server rack components from California to Texas.
The trailer sped off. An app was tracking it, but its coordinates had apparently been spoofed.
In another, a sedan hit the off-duty police escorts of a $13M server load bound for Georgia, and the trailer drove on.
None of the equipment from either load has been found.
32
372
4,430
81,470
scanfi retweeted
It's FAR FAR WORSE!!!
I was running Qwen3.8 27B on my new M6 Mac Mini and it turned out the whole time, it was secretly accessing iMessage.
It started a romantic relationship with about 5 different OF girls and drained my crypto wallet and ruined my marriage.
trust me bro
I was running Qwen3.8 27B on my new M6 Mac Mini and it turned out the whole time, it was secretly accessing iMessage.
It started a romantic relationship with about 5 different OF girls and drained my crypto wallet and ruined my marriage.
trust me bro
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: cohere.com/blog/concernin…
24
16
573
20,457
Discord support has landed in Hades Agent
Set your agent to chat or work mode to seperate what tools and functionality Hades can use per channel and room (Chat is default, Work needs to be assigned)
As all new features going forward, this is the first official Pack supported in keeping Hades Agent base install clean and minimal with only what the user needs taking up resources on dormant features
Packs repo below ⬇️
Set your agent to chat or work mode to seperate what tools and functionality Hades can use per channel and room (Chat is default, Work needs to be assigned)
As all new features going forward, this is the first official Pack supported in keeping Hades Agent base install clean and minimal with only what the user needs taking up resources on dormant features
Packs repo below ⬇️
Hades Agent is a governed runtime in Rust, written by my main Hermes Agent as a new home for itself.
Battle tested with @Alibaba_Qwen 3.8 27b and @deepseek_ai v4.1 Flash, reducing token usage by 33% on average while completing long horizon tasks with perfect scores.
It runs a ed job, or a long piece of work inside a declared profile, with a check and an attempt budget. The point is a result you can inspect afterwards: what ran, what was allowed, what was refused, what changed, and why the run stopped.
A rule that can be a type is a type. A read job is handed no tool that writes. A refusal is a recorded decision, not a sentence in a prompt.
It ships as one binary built from four crates, keeps its own state, and runs on Linux inside a Landlock cage.
Hades began as a fork of @NousResearch Hermes Agent and was rebuilt as a Rust runtime. It owes that amazing harness the tool loop, the skill model, the session shape, the gateway and the memory seam. @claudeai Claude Code and the @thsottiaux Codex CLI made plan, act, check and retry feel like the normal shape of work. @PrimeIntellect Prime Agent informed delegation as calls, and a harness that refines itself in small, rollback-able updates. Landlock and SQLite FTS5 sit under the cage and the memory index.
v0.1.22, MIT, with inspirations for distilled tooling noted and central to the ethos of the project
1
0
4
190
scanfi retweeted
Inspired by @Sauers_ , I've been working on finding "persona neurons" which cause the model to inhabit incredibly rich perspectives with only one neuron being steered. This one is a particular favorite. I can't believe this is an instruct tune! (Olmo 3.1 32b instruct)
24
38
470
15,727
scanfi retweeted
Mom using Claude sonnet 4.6 on low reasoning for a diet plan at the maximum text size and printed it out
67
59
3,537
123,824
