HN Debrief

Ornith-1.5: From Self-Scaffolding to Self-Improvement

  • AI
  • Open Source
  • Developer Tools
  • Hardware

Ornith released a new family of open-weight models and framed the work as moving beyond agent wrappers that help a model reason into training loops that improve the model itself. That headline did not persuade many people on its own. Most readers quickly reduced the claim to a more concrete question: is this just Qwen post-training plus a harness, and does the resulting 35B mixture-of-experts model earn a place on a local machine? The answer from the comments was pragmatic. Yes, these appear to be post-trained Qwen models, not a new base model from scratch, and the interesting part is whether the post-training materially changed the tradeoff between speed, memory, and coding quality.

If you care about local coding agents, treat Ornith-1.5 as a promising fork of Qwen rather than a breakthrough in autonomous model improvement. Test it on your own prompts and hardware, because quantization, context length, and memory layout seem to matter more here than benchmark tables or the marketing language.

Discussion mood

Interested but skeptical. People liked having another open-weight local model option and several early users reported good speed and decent tool use, but the bigger mood was distrust of the self-improvement framing and uncertainty that Ornith-1.5 actually surpasses Qwen 3.8 on coding once you control for quantization and hardware.

Key insights

  1. 01

    This is post-trained Qwen, not a new base

    The key fact missing from the announcement is that these models appear to be built by post-training existing Qwen checkpoints. That sharply narrows what the release proves. It is evidence that Ornith's training harness can move strong open models around the frontier. It is not evidence they trained a competitive base model from scratch or discovered a fundamentally new architecture.

    Read the results as a test of post-training quality and deployment fit. If you are evaluating Ornith for use or investment, ask what data, reinforcement loop, and tooling changed the inherited Qwen behavior enough to matter.

      Attribution:
    • goldemerald #1
    • kees99 #1
  2. 02

    Public benchmarks are weak buying guides

    People with hands-on local setups were blunt that benchmark tables do not decide model choice anymore. Public evals are easy to optimize against and often miss the details that wreck a real workflow, especially coding agents, scraping, or long-context tool use. The only numbers that carry weight are the ones built from your own tasks and kept private enough that model makers cannot train to them.

    Build a small internal eval set around your actual tasks before standardizing on a local model. Include your runtime, quant, context size, and tool loop, because those choices can change the winner.

      Attribution:
    • swatcoder #1
    • NitpickLawyer #1
  3. 03

    MoE helps only in certain local setups

    The useful framing is not that mixture-of-experts is simply faster or simply worse. It wins when you can store the model but cannot afford to activate a dense network every token. It loses when VRAM capacity is already the limiting factor and a dense model with the same memory footprint gives you more intelligence. That explains why some people swear by 35B-A3B on 16 GB cards while others have moved back to dense 27B models on stronger boxes.

    Match architecture to hardware before comparing model labels. On constrained consumer GPUs, test both a compact mixture-of-experts model and a denser model at the highest quant you can tolerate instead of assuming one class dominates.

      Attribution:
    • c0m47053 #1
    • verdverm #1
    • Sha1rholder #1
  4. 04

    Quantization and runtime can flip the verdict

    Several comments make clear that people are often comparing different quants, runtimes, and decoding setups more than different models. One person saw Qwen 3.8 27B as a huge coding upgrade. Another found it slower and worse than an older Qwen mixture-of-experts model until runtime choices like Lightning MTP entered the picture. A third point cut through the confusion: it is easy to make a good model look dumb by over-quantizing it or copying settings from a different model family.

    When someone claims a model is better or worse, ask for the full inference stack. Keep a reproducible matrix of quant, runtime, decoding method, and prompt settings, or you will chase phantom regressions.

      Attribution:
    • smcleod #1
    • xscott #1
    • seanmcdirmid #1
  5. 05

    Long context usability matters more than peak IQ

    What sold some early users was not a claim that Ornith-1.5 is the smartest model. It was that they could hold 32k context at workable token speeds on Apple silicon, which made agentic coding practical instead of frustrating. That is a different win from benchmark leadership and one that better matches how local assistants are actually used.

    For coding agents and research assistants, measure sustained performance at the context lengths you really need. A slightly weaker model that stays fast and stable at 32k can beat a stronger one that becomes unusable there.

      Attribution:
    • huseyinkeles #1
    • quinncom #1
    • smcleod #1
  6. 06

    Community distillation is possible but expensive

    One commenter asked the obvious question: if people want a Qwen 3.8-style 35B-A3B, why not distill it from the giant open model? The answer was yes in principle, and hobbyists have already posted unofficial distilled variants, but the bottleneck is training cost. Renting enough high-end hardware to do a serious distillation run is exactly why there is demand for official small models in the first place.

    Do not assume open weights automatically produce the exact model tier the market wants. There is still room for startups that turn big released checkpoints into efficient, trustworthy smaller models.

      Attribution:
    • WASDx #1
    • halJordan #1
    • boznz #1

Against the grain

  1. 01

    Qwen 3.8 27B still looks stronger

    A direct comparison against published coding and reasoning scores pushed back on the enthusiasm around Ornith-1.5-35B. On the cited table, Qwen 3.8 27B was ahead on Terminal-Bench, SWE-bench Pro, DeepSWE, and Humanity's Last Exam, with Ornith only leading on NL2Repo and tying GPQA Diamond. That makes Ornith look less like a new leader and more like a useful deployment tradeoff.

    If you can fit and run Qwen 3.8 27B well, do not assume Ornith-1.5-35B is an upgrade. Start from the hypothesis that it is a speed and packaging play, then see if your workflow says otherwise.

      Attribution:
    • hgoel #1
    • ricardobeat #1
  2. 02

    The RSI story reads like fundraising theater

    The harshest read was that the post leans on vague recursive self-improvement language to sound frontier-grade without proving much. That skepticism changes how to interpret the announcement. The safest reading is not a breakthrough claim about autonomous model improvement. It is a startup marketing a competent model release in the language investors currently want to hear.

    Separate the product from the narrative. Evaluate the shipped weights and the reproducible gains, not the grand theory attached to them.

      Attribution:
    • htrp #1
  3. 03

    Identity confusion is not meaningful evidence

    One early complaint was that the model kept claiming to be Claude. Others dismissed that as nearly universal behavior in open models rather than a special tell about Ornith. Model identity is often shallowly trained, can vary by language, and says little about actual capability or provenance unless the failure is unusually persistent in a controlled test.

    Do not overweight persona glitches when screening open models. Focus first on task behavior, tool use, and failure modes that affect production work.

      Attribution:
    • wgd #1 #2
    • orbital-decay #1
    • garo-pro #1

In plain english

agentic coding ↩
Using AI systems that can plan and carry out multi-step software development tasks with limited human intervention.
Apple Silicon ↩
Apple’s custom ARM-based chips used in modern Macs, such as the M1, M2, M3, and M4 families.
DeepSWE ↩
A benchmark focused on software engineering tasks for AI coding systems.
distill ↩
Train a smaller model to imitate a larger one so it keeps some of the larger model’s behavior at lower cost.
GPQA Diamond ↩
A hard question-answering benchmark used to test advanced reasoning on graduate-level science questions.
MTP ↩
Multi-token prediction, a speculative decoding method where a model predicts multiple future tokens to increase generation speed.
NL2Repo ↩
A benchmark that measures how well a model maps natural language requests to changes or actions in a code repository.
SWE-bench Pro ↩
A benchmark that tests whether models can solve real software engineering tasks in codebases.
Terminal-Bench ↩
A benchmark that evaluates how well models perform tasks through terminal-style interactions.
VRAM ↩
Video random-access memory, the memory attached to a graphics processor.

Reference links

Model releases and collections

Tools for local inference

  • samosa-chat
    An open source local chat tool someone used to run Ornith-1 locally.

Background on company and founder