[HN Gopher] AI2: Open Coding Agents
       ___________________________________________________________________
        
       AI2: Open Coding Agents
        
       Author : publicmatt
       Score  : 92 points
       Date   : 2026-01-27 17:17 UTC (5 hours ago)
        
 (HTM) web link (allenai.org)
 (TXT) w3m dump (allenai.org)
        
       | jauntywundrkind wrote:
       | Awesome stuff. Output speed looks crazy fast too.
       | 
       | I wonder if this indeed will start prompting more language
       | specific work.
       | 
       | Afaik training still requires not just looking at sample code but
       | also being able to write loss functions being able to have
       | problems the AI can work at. That seems hard.
       | 
       | One random thought, are there training styles of just deleting
       | some code from "good" projects then making the AI make it work
       | again?
        
         | CuriouslyC wrote:
         | The technique people use is to capture PR diffs from public
         | repos and extract the tests then use that to see if agents can
         | reconstruct the patch that satisfies the tests.
        
       | ahmadyan wrote:
       | Claims in the article are incorrect. They conveniently ignore
       | Meta CWM models, which are open-sourced [1] and open-weight [2]
       | and are at 65% SWE-bench verified (with TTS) and 54% pass@1 and
       | the same size (32B dense). So claims like "surpassing prior open-
       | source state-of-the-art coding models of comparable sizes and
       | context lengths" and conveniently leaving out the previous OSS
       | SOTA out of your eval tables are ... sketch.
       | 
       | [1]https://github.com/facebookresearch/cwm
       | [2]https://huggingface.co/facebook/cwm
        
         | philipkglass wrote:
         | The difference is that the Allen Institute models have open
         | training data, not just open code and weights. Meta doesn't
         | share the training data you would need to reproduce their final
         | models. For many uses open-weight models are nearly as good,
         | but for advancing research it's much better to have everything
         | in the open.
        
           | kevmo314 wrote:
           | Reading their paper, it wasn't trained from scratch, it's a
           | fine tune of a Qwen3-32B model. I think this approach is
           | correct, but it does mean that only a subset of the training
           | data is really open.
        
         | mhitza wrote:
         | The linked open weight disallows commercial, and is only
         | licensed for research purpose
        
         | ethan_l_shen wrote:
         | Hey! These are great observations. So first, while TTS can
         | improve performance, we wanted to evaluate the raw capability
         | of our model. This meant generating only one rollout per
         | evaluation instance, which follows other papers in the space
         | like SWE-smith and BugPilot. In addition, TTS adds extra
         | inference cost and is reliant on how rollouts are ranked, two
         | confounding factors for deployable models where memory and
         | inference speed are extremely important.
         | 
         | Following that line of reasoning, context length is another
         | very large confounding factor. Longer context lengths improve
         | performance - but also result in enormous increases in KV cache
         | size and memory requirements. We decide to control for this in
         | our paper and focus at the 32K context length for 32B size
         | models, a context length that already pushes the bounds of what
         | can be "deployable" locally.
         | 
         | Still, we evaluate at 64K context length using YARN and are
         | able to outperform CWM's 54% performance (non TTS), which it
         | achieves using 128K context, a substantial increase over what
         | we use. This is also pretty significant because we only ever
         | train at 32K context, but CWM trains for a full 128K.
        
       | khimaros wrote:
       | it's great to see this kind of progress in reproducible weights,
       | but color me confused. this claims to be better and smaller than
       | Devstral-Small-2-24B, while clocking in at 32B (larger) and
       | scoring more poorly?
        
         | ethan_l_shen wrote:
         | Hey! We are able to outperform Devstral-Small-2-24B when
         | specializing on repositories, and come well within the range of
         | uncertainty with our best SERA-32B model. That being said, our
         | model is a bit larger than Devstral 24B. Could you point out
         | what in the paper gave the impression that we were smaller? If
         | theres something unclear we would love to revise
        
           | khimaros wrote:
           | "SERA-32B is the first model in Ai2's Open Coding Agents
           | series. It is a state-of-the-art open-source coding agent
           | that achieves 49.5% on SWE-bench Verified, matching the
           | performance of much larger models like Devstral-Small-2
           | (24B)" from https://huggingface.co/allenai/SERA-32B
        
             | ethan_l_shen wrote:
             | Ah great catch I don't know how we missed that. Thanks!
             | Will fix.
        
       | augusteo wrote:
       | The ahmadyan comparison is fair. Meta's CWM models hitting 65% vs
       | SERA's 54% is a meaningful gap.
       | 
       | But the interesting number here isn't accuracy. It's the $400 to
       | reproduce top open-source performance. That's the part that
       | matters for teams building internal tooling.
       | 
       | We've been running agents on proprietary codebases at work. The
       | pain isn't model quality. It's customization. Most off-the-shelf
       | agents don't understand your repo structure, your conventions,
       | your test patterns. If you can fine-tune a 32B model on your own
       | codebase for a few hundred dollars, that changes the economics
       | completely.
       | 
       | But codebases changes everyday, so finetuning will have to be
       | continuously done!
       | 
       | Probably not worth it versus something like Claude Code.
       | 
       | Curious whether anyone's tried this on non-Python codebases. Most
       | SWE-Bench stuff is Python-heavy.
        
         | storystarling wrote:
         | The fine-tuning overhead is definitely a factor, but for
         | smaller shops the hard constraint is usually inference VRAM.
         | Running a 32B model locally or on a rented GPU is surprisingly
         | expensive if you aren't saturating it. Even at 4-bit
         | quantization you are looking at dual 3090s or an A6000 to get
         | decent tokens per second. The $400 training cost is impressive
         | but the hosting bill is what actually kills the margin compared
         | to per-token APIs.
        
       | nickandbro wrote:
       | Great work! Really respect AI2. they open source everything. The
       | model, the weights, the training pipeline, inference stack, and
       | corpus
        
       | Imustaskforhelp wrote:
       | Hey this looks great? Is it available on Openrouter.
       | 
       | I wish if AI2 could release a more denser model on Openrouter for
       | free than the 8B model as I was using Devstral model for agentic
       | purposes.
       | 
       | If we can get an agentic good 32B like model on openrouter for
       | ~free, then I feel like it will be very interesting to see how
       | things would go imo.
       | 
       | Good luck with AI2! The premise of truly open source models is
       | really interesting and I feel like it could help bring more
       | innovation in the space imo!
        
       | ripped_britches wrote:
       | One claim in article is definitely very wrong or at least needs
       | to be narrowed. Claude is the only closed agent harness and there
       | are about two dozen open ones. Many models may be closed, but
       | when people say agent they are generally referring to the
       | harness, not the underlying model.
        
       | janmue wrote:
       | "Strong closed-weight coding agents like Devstral Small 2 are an
       | important point of comparison."
       | 
       | Devstral Small 2 is an open-weights model:
       | https://huggingface.co/mistralai/Devstral-Small-2-24B-Instru...
        
       ___________________________________________________________________
       (page generated 2026-01-27 23:00 UTC)