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.)
> 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.
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 ?
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.
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.
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.
> 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?
> 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.
Why is Qwen3.5 2B not in the table?
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.