Rendered at 23:43:41 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
vezycash 14 days ago [-]
Always put "Lower is better" or "higher is better" in benchmarks. Not everyone knows what your numbers mean.
bbminner 13 days ago [-]
Actually, I came to comments specifically to praise the inverted bar charts, I've never seen it before, a very intuitive way to communicate lower is better IMHO.
mytailorisrich 13 days ago [-]
I'd say not really because most common use is to plot metrics like deficits where lower is worse so people may be accustomed to that... thus best practice should indeed be to always label whatever the direction of the chart to make clear what is better or worse.
leothetechguy 13 days ago [-]
If a score is better when it's low, isn't that measuring a deficit? I think it's intuitive.
lend000 14 days ago [-]
I can't imagine who is using something like this for agentic coding, but I see exciting opportunities on the horizon when we can have hundreds of reasonably rational and conversational agents working on local machines to simulate emergent behavior (simulating crowds, markets, ecosystems, game NPCs, etc.)
trvz 14 days ago [-]
Agentic behaviour isn't just coding. In fact:
> We recommend using it for agentic workloads, tool use, data extraction, RAG, and long-context workflows. It is not recommended for agentic coding and knowledge-heavy tasks.
touisteur 14 days ago [-]
Really curious about people's workflows with these agentic-but-not-for-coding workflows. Are there some interesting people to follow there or just good testbeds/environments to get an idea ?
dd8601fn 14 days ago [-]
My little home assistant.
It has to query the tool service, invoke tools, synthesize results, or request new tools. Nothing really complex.
New tool requests from it are plain english and go into a separate pipeline using more appropriate models. It doesn’t have to write anything itself.
dofm 14 days ago [-]
Yeah. Clearly the target for these models is on-device assistant/analyst tasks where really everything is agentic.
Thinking out loud: any text-based generative AI application is on a continuum between:
- document-to-document: solve this language problem using language skills by outputting only new symbols, and
- document-to-tasks: solve this language problem by only operating these tools.
In a sense, given access to vast compute to train a very large model, agentic coding is one of the easier applications that is somewhere between the two. It is manipulating symbols in a mixture of languages that are biased towards context-free (code), it exploits the embedded knowledge in a vast number of weights, and it calls fairly simple tools. The user-focussed solution is happening inside the LLM.
A small model can't do this job well, even if it has a good understanding of programming languages, because it lacks the world knowledge to understand the problem.
At the other end of the continuum is: these words mean do this thing. That Cactus Needle 2 model mentioned earlier is here. It doesn't even produce a language description of what it did, because it one-shots tasks.
In the middle but near this end is the fantastically hard job: train a model that understands language and reasons well enough to respond to queries about the tools it has access to, operate them and reply in natural language, without being large or slow. That is, understand language without being overburdened by details of scenarios that caused the need for the words in the first place.
This is maybe the great grand challenge of LLMs: make it know how to speak and reason and think and work for us while discarding everything that is just knowledge of unnecessary facts.
Small reasoning models are where most of the really big challenges are.
jwr 13 days ago [-]
Removing spam from email, extracting attachments (invoices) from emails, and otherwise managing E-mail. Those are my use cases, although I suspect this model will be too small for that. I currently use 27-35B models (MoE mostly).
yieldcrv 14 days ago [-]
I feed these things my tax transcripts and it found all sorts of refunds and ways to waive penalties that the IRS and my state agreed I was eligible for, because my current CPA doesn’t do this kind of tax matter
The high parameter 2026 models know the latest tax law, revealed deadlines for court rulings to me that are in fact real and new branches of service on the IRS website, or can learn while also agentically search
While many debate the utility in legal matters and 2023’s hallucination issue, the rest of us just do
dannyw 13 days ago [-]
So many people have outdated notions of what LLMs can do; and tbh, I think a lot of it comes down to free models being weaker.
jurgenburgen 14 days ago [-]
Using climate-destroying AI to understand the overly convoluted tax regime in the US is peak capitalism.
Yiin 13 days ago [-]
I'd start with your fridge and ac
13 days ago [-]
perching_aix 13 days ago [-]
Climate destroying? How do they destroy the climate by hosting a few gigabyte model on their computer?
It's the whole point with these. Uses hardware they already own, uses energy they very possibly produce for themselves.
Or is this just me needing to take a joke?
trvz 13 days ago [-]
“High parameter […] models” was mentioned above. It’s still not a good take however, as laws of civilizations getting more convoluted over time is the general behaviour, not specific to LLM-producing countries.
l3x4ur1n 14 days ago [-]
I am getting to know Hermes agent and in the last couple of days I used it for searching the web for medical specialists or finding the best deal for a product online, manipulate with excel and word documents, gather data from APIs, finding stuff in my computer or keeping track of itself running in another process and keep the computer awake and when detecting the process is finished, put computer to sleep.
htrp 13 days ago [-]
Better question is does a coding model do better on these agentic tasks? And is the only advantage efficiency??
antupis 14 days ago [-]
I have noticed that these cheaper and faster models are very great for Ops-work. Luna max is beast when you use some stronger model to write detailed instructions/run book what to do and when to stop.
jofzar 13 days ago [-]
I use luna to schedule my work day for me and then put it in todoist.
theblazehen 13 days ago [-]
I use them for game AI
PcChip 13 days ago [-]
have any details, writups on a blog, or videos?
sounds interesting!
theblazehen 13 days ago [-]
Still a work in progress, but planning on publishing a post when it's complete
snek_case 13 days ago [-]
You can use it for things like classifying documents, spam detection and detecting certain events, etc. You're right that it's not strong enough for coding but it can still be quite good for question answering and such. Or even simple things like converting documents from one format to another.
And yeah, for NPCs, this model and their other A1B model can fit inside of a game and run at a decent speed on CPU-only. It would probably be workable for game characters.
0xbadcafebee 14 days ago [-]
LFM's training/post-training is famously different than other models. They target reliable operation of tiny models in ways other model families don't (they aren't just scaling a larger model to a smaller size). If you're looking for good performance out of tiny models, LFM has the most advanced design.
Note how they're much smaller than all other models in the comparison yet match or exceed them. This is for 2.6B params, but they have models as small as 230M. Nobody else designs models that small.
woadwarrior01 14 days ago [-]
> Note how they're much smaller than all other models in the comparison yet match or exceed them.
There's a strong incentive to cherry pick in self-reported comparisons. If there is a model that's better, it gets left out. Have you seen Nanbeige4.2-3B or Ling-3.0-tiny?
> Nobody else designs models that small.
There are people building even smaller models.
jhgik798 13 days ago [-]
Granite IBM make tiny
h14h 13 days ago [-]
I love these smaller models for the added capability they can bring to older hardware.
Recently picked up a 2013 Mac Pro (the trashcan one) on eBay for $150, and the official 6-bit quant of 2.6B runs at 25-30 tok/s under llama.cpp on one of the D500 GPUs. After installing Ubuntu LTS I have myself a lovely (if somewhat inefficient) little homelab server.
Gecko4072 14 days ago [-]
These LiquidAI models have never worked well for me in practice.
eurekin 14 days ago [-]
Care to share any details? I'm about to check the 2.6b lfm on document editing.
weebull 13 days ago [-]
I found that they didn't want to call tools in the same way as other models, and it led to bare tool calls in the response.
NitpickLawyer 13 days ago [-]
How are you serving them? This is most often caused by an incorrect template. (the thing that tells the inference engine how to parse the think/tool parts of the answer)
For vllm the official recipe [1] (for another model in the same family) is this:
> We recommend using it for agentic workloads, tool use, data extraction, RAG, and long-context workflows. It is not recommended for agentic coding and knowledge-heavy tasks.
It has to query the tool service, invoke tools, synthesize results, or request new tools. Nothing really complex.
New tool requests from it are plain english and go into a separate pipeline using more appropriate models. It doesn’t have to write anything itself.
Thinking out loud: any text-based generative AI application is on a continuum between:
- document-to-document: solve this language problem using language skills by outputting only new symbols, and
- document-to-tasks: solve this language problem by only operating these tools.
In a sense, given access to vast compute to train a very large model, agentic coding is one of the easier applications that is somewhere between the two. It is manipulating symbols in a mixture of languages that are biased towards context-free (code), it exploits the embedded knowledge in a vast number of weights, and it calls fairly simple tools. The user-focussed solution is happening inside the LLM.
A small model can't do this job well, even if it has a good understanding of programming languages, because it lacks the world knowledge to understand the problem.
At the other end of the continuum is: these words mean do this thing. That Cactus Needle 2 model mentioned earlier is here. It doesn't even produce a language description of what it did, because it one-shots tasks.
In the middle but near this end is the fantastically hard job: train a model that understands language and reasons well enough to respond to queries about the tools it has access to, operate them and reply in natural language, without being large or slow. That is, understand language without being overburdened by details of scenarios that caused the need for the words in the first place.
This is maybe the great grand challenge of LLMs: make it know how to speak and reason and think and work for us while discarding everything that is just knowledge of unnecessary facts.
Small reasoning models are where most of the really big challenges are.
The high parameter 2026 models know the latest tax law, revealed deadlines for court rulings to me that are in fact real and new branches of service on the IRS website, or can learn while also agentically search
While many debate the utility in legal matters and 2023’s hallucination issue, the rest of us just do
It's the whole point with these. Uses hardware they already own, uses energy they very possibly produce for themselves.
Or is this just me needing to take a joke?
sounds interesting!
And yeah, for NPCs, this model and their other A1B model can fit inside of a game and run at a decent speed on CPU-only. It would probably be workable for game characters.
Note how they're much smaller than all other models in the comparison yet match or exceed them. This is for 2.6B params, but they have models as small as 230M. Nobody else designs models that small.
There's a strong incentive to cherry pick in self-reported comparisons. If there is a model that's better, it gets left out. Have you seen Nanbeige4.2-3B or Ling-3.0-tiny?
> Nobody else designs models that small.
There are people building even smaller models.
Recently picked up a 2013 Mac Pro (the trashcan one) on eBay for $150, and the official 6-bit quant of 2.6B runs at 25-30 tok/s under llama.cpp on one of the D500 GPUs. After installing Ubuntu LTS I have myself a lovely (if somewhat inefficient) little homelab server.
For vllm the official recipe [1] (for another model in the same family) is this:
[1] - https://recipes.vllm.ai/LiquidAI/LFM2.5-8B-A1BWhy is Qwen3.5 2B not in the table?