https://swelljoe.com/post/will-it-mythos/: "Poor performer here, only found the one bug that almost every model found, despite its performance on other benchmarks being excellent for its size. […] It also performs poorly in a chat without tools, exhibiting an ehthusiasm for hallucination. I’m currently working on a replication of this with full tool access, including bash/Python, which may allow this model to be competitive."
How is that a serious phrase in '26? I mean I have no idea if this fine-tune is good, haven't tried it, but testing a (clearly) agentic model without tool access and expecting it to work is crazy, no? What was he even testing?!
It has been this way since the beginning, unfortunately. There is certainly no harm in trying on local models on local workloads with modest guardrails.
Like most of these models (Qwen, Gemma, Llama, gpt-oss), finding all the little gotchas like, special tokens and prompt structure, model preference are a PITA right now. The reward are really nice models that run exceptionally well in agentic harnesses tuned with the prompts and parameters you fought so hard to learn.
It gave a fancier looking answer, but did a worse job following the prompt.
Ah, the place that shit on gpt-oss because it wasn't good at porn. That place is not what it used to be, hasn't been since that karpathy tweet, tbh. It's mostly slop and vibes nowadays.
I use it via llamacpp and codex-cli.
How does it self-improve, does the model change on disk - or just during a single context run it gets better?
As far as I can tell they trained it by running their own reinforcement learning on top of Qwen and Gemma 4 (not sure how they combined weights from both, or if they used Qwen as the basis and Gemma 4 to help train?) - so the "self-improving" is about their training process, not how you use the weights.
It does good job on basic analysis, tasks, and some front-end/backend changes on a medium-sized Go codebase, but it reached its limits totally botching a longer (simple) kernel implementation job (about 100 iterations in Pi Agent harness) - this is the type of thing that stronger open models (Kimi K2.6, GLM 5.2) are able to do.
Model Adapter Suite Score Passed Tasks
--------------------------------- ------------ -------------- ------ ------ -----
local/ornith-1.0-35b little_coder aider_polyglot 36.0% 81/225 225
local/ornith-1.0-35b pi_devstack aider_polyglot 39.6% 89/225 225
local/ornith-1.0-35b pi_vanilla aider_polyglot 32.0% 72/225 225
>Built on top of pretrained Gemma 4 and Qwen 3.5, it achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks.
>Ornith-1.0 is a self-improving training framework. Instead of relying on human-designed harnesses to drive solution generation in RL, Ornith-1.0 learns to generate both solution rollouts and the task-specific harnesses that guide those rollouts.
Us mere mortals cannot use this.
However, it's much more inclined to do web search unprompted, which is fascinating in its own way.
All the rest - including 9B - are Qwen 3.5/3.6:
https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B/blob/m...