Ornith-1.0 Deep Dive — DeepReinforce's June 26, 2026 MIT Open-Weights Family Specialized for Agentic Coding Three Sizes (9B Dense / 35B MoE / 397B MoE), All at 262K Context, Built on Qwen 3.5 + Gemma 4, Shipping in BF16 + FP8 + GGUF SWE-Bench Verified 82.4% (397B) / 75.6% (35B) / 69.4% (9B), SWE-Bench Pro 62.2%, Vendor-Reported SOTA Among Open Weights at Each Size Tier Reinforcement Learning Optimizes Both Solution Rollouts AND the Scaffolding That Drives Them — A 'Self-Improving' Design Compatible With OpenHands / Hermes Agent / OpenClaw, ClawEval Benchmark Published — Directly Relevant to Oflight's OpenClaw Service Users

DeepReinforce released Ornith-1.0 on June 26, 2026 (official / Hugging Face collection).

Four points:

1. MIT open-weights LLM family specialized for agentic coding — no regional restrictions 2. Three sizes ship simultaneously — 9B Dense / 35B MoE / 397B MoE, all at 262K context, with FP8 and GGUF quantizations released alongside 3. SWE-Bench Verified 82.4% (397B) / 75.6% (35B) / 69.4% (9B), SWE-Bench Pro 62.2%, Terminal-Bench 2.1 77.5-78.2% — claimed SOTA at each open-weights size tier 4. Design signature: RL optimizes both solution rollouts AND the scaffolding that drives them — a self-improving design, with first-class compatibility for OpenHands / Hermes Agent / OpenClaw (ClawEval is in DeepReinforce's own benchmark suite)

This column sits next to our Kimi K2.7-Code, Local LLM June 2026 Update, and Loop Engineering coverage as part of the June-2026 agentic-coding cluster.

DeepReinforce is an RL-focused research organization that has previously published:

- [CUDA-L1](https://github.com/deepreinforce-ai/CUDA-L1): an RL framework for CUDA optimization, with avg 3.12× speedup across 250 real-world GPU tasks (MarkTechPost, 2025-08) - [CUDA-L2](https://github.com/deepreinforce-ai/CUDA-L2): RL-based kernels surpassing cuBLAS on matrix multiply (RTX 3090 HGEMM in March 2026, A100 HGEMM in January 2026) - IterX: presented at MLSys 2026 NVIDIA Track with significant H100 / B200 speedups - Ornith-1.0: applies the same RL playbook to LLM self-improvement

Reading Ornith-1.0 as the next step in this RL-everywhere research line is the right frame.

The headline design choice in Ornith-1.0 is that reinforcement learning optimizes both the solution rollouts and the scaffolding (the agent structure that drives those rollouts) — not just one or the other.

Standard RLHF / RLAIF optimizes input → output in one shot. Ornith-1.0 instead trains over input → the agent's reasoning trajectory and tool-call sequence → final output as a single end-to-end RL objective. The consequences:

- The model itself learns how to decompose a problem - The decision policy for when to call which tool gets baked into the weights - Running it inside an agent framework (OpenHands / Hermes / OpenClaw), the scaffolding behavior is already learned, so fewer trials are needed to reach a correct answer

This is essentially the Loop Engineering Maker-Checker pattern internalized into the training objective. It's a different lineage from Sakana Fugu's orchestration model (which dispatches across multiple LLMs); Ornith learns the scaffolding inside a single LLM.

What stands out:

- 397B at SWE-Bench Verified 82.4% — open-weights at that level is in the same neighborhood as Kimi K2.7-Code (which didn't publish SWE-Bench, only vendor-internal benches) - 35B at SWE-Bench Verified 75.6% — within striking distance of Claude Opus 4.8 / GPT-5.5, and runnable on a serious consumer GPU stack (e.g. 8×RTX 5090) - 9B Dense at SWE-Bench Verified 69.4% — a frontier-class agent on a single 80GB GPU or one RTX 5090 - ClawEval 77.1% on 397B — OpenClaw is Oflight's own service line, and seeing DeepReinforce treat ClawEval as a first-class benchmark is meaningful industry validation

Important caveat: these are DeepReinforce's own self-reported numbers. Registration on the public SWE-Bench leaderboard and third-party validations on Aider polyglot / LiveCodeBench / Cognition FrontierCode have not appeared as of June 26, 2026. Run a PoC on your own code before production adoption.

Ornith-1.0-9B (Dense):

- 9B dense transformer - 262K context, BF16 - Single 80GB GPU (~19GB VRAM) or comfortable on one RTX 5090 (32GB) - Tensor parallelism supports multi-GPU sharding - Realistic production candidate for individual developers and SMBs

Ornith-1.0-35B (MoE):

- 35B Mixture-of-Experts - Built on Qwen 3.5 - 262K context, BF16 - Recommended: 8×80GB GPU node (TP=8); the GGUF quant lets you go lighter - DeepReinforce reports beating Qwen 3.5-35B and Gemma 4-31B on benchmarks - The mid-market production sweet spot

Ornith-1.0-397B (MoE):

- 397B Mixture-of-Experts - Built on Qwen 3.5 + Gemma 4 (hybrid composition) - 262K context, BF16 (FP8 quant available) - Recommended: 8×80GB GPU node (TP=8) - SWE-Bench Verified 82.4% / SWE-Bench Pro 62.2% / Terminal-Bench 2.1 77.5-78.2% — claimed open-weights SOTA at this size tier - The flagship for large-enterprise SI

Recommended inference engines:

- vLLM ≥ 0.19.1 — production GPU serving - SGLang ≥ 0.5.9 — agent workflows (RadixAttention) - Transformers ≥ 5.8.1 — Hugging Face standard - Docker + llama.cpp — CPU / edge via GGUF - Ollama — personal PoC

API compatibility: OpenAI-compatible endpoints (via vLLM / SGLang). Drops into existing tools (Claude Code, Cursor, Aider, Cline, cmux) with config-only changes.

Ornith-1.0 is officially compatible with:

- OpenHands - Hermes Agent (Nous Research) - [OpenClaw](../services/openclaw-setup) (Oflight's own service line; ClawEval is in DeepReinforce's published benchmarks)

Impact for OpenClaw users: customers running our OpenClaw setup service can swap Ornith-1.0 in as the backend LLM with low effort — change the API key and inference endpoint and you get ClawEval-77.1%-class behavior on the 397B.

Our OpenClaw monthly maintenance plans (Light ¥9,800 / Standard ¥19,800 / Premium ¥49,800) include LLM-model swaps and API-spec updates — Ornith-1.0 migration falls within scope.

Ornith-1.0's differentiators:

1. Full three-size MIT open-weights line — same family from individual workstation (9B) to enterprise flagship (397B) 2. US (presumed) origin with no regional restrictions — sidesteps the Chinese-model cross-border-data scrutiny 3. Published benchmarks include the standard SWE-Bench suite — easier to compare than Kimi K2.7-Code's vendor-only numbers 4. RL-optimized scaffolding — a real-world agent-runtime edge by design 5. ClawEval treated as first-class — strong fit for OpenClaw adopters

- Large refactors and multi-file PRs (35B / 397B) - Automated code review inside CI / CD (9B / 35B) - Agent-style SWE-Bench-class problems (all sizes) - Terminal agents via Cline / Aider / cmux (35B recommended) - Multi-turn tool-use workflows - Agentic coding on OpenClaw / OpenHands / Hermes - On-prem, data-sovereignty-sensitive SI engagements

What we recommend in our AI consulting practice:

Step 1 — PoC on the 9B (from ¥198K, AI consulting assessment): 9B fits on a single RTX 5090 / 80GB GPU, so PoC hardware cost is minimal. Measure ROI on your own coding workloads (CI auto-review, internal SDK Q&A, bug triage).

Step 2 — Production on 35B / 397B (from ¥498K PoC build, ~¥5M+ full SI): 35B fits on a single 8×H100 / B200 node; 397B wants B200 ×4 / H200 ×8 class. Pair with a Japanese GPU cloud (Sakura HPC, GMO GPU, AWS Tokyo p5) or the Intec ¥5M+ SI pattern.

Step 3 — OpenClaw + Ornith for agent-based business automation (with [OpenClaw maintenance plan](../services/openclaw-setup) ¥9.8K–¥49.8K): customers already on OpenClaw can swap Ornith-1.0 in as the backend LLM and pick up ClawEval-77.1%-class performance. LLM swap and prompt tuning are in scope of OpenClaw monthly maintenance.

As of June 26, 2026, the following were not confirmable in our research:

- DeepReinforce's HQ location, executive team, and funding status (the official deep-reinforce.com returned 403 to our automated fetcher, so this was not confirmable via the web) - Hosted SaaS API availability and pricing (the product may be self-host / on-prem only) - A Japan-region endpoint - Plans for registration on the [public SWE-Bench leaderboard](https://www.swebench.com/) - Third-party scores on Cognition FrontierCode, Aider polyglot, LiveCodeBench, BigCodeBench - Official VS Code / JetBrains / IntelliJ extensions - AIME / GPQA-Diamond / MMLU and other general-knowledge benchmarks (potentially omitted on purpose, as the model is coding-specialized)

Verify with the Hugging Face collection and DeepReinforce GitHub before any production decision.

We support Ornith-1.0 and other local LLMs across evaluation, build, and ongoing maintenance.

"Does Ornith-1.0 fit our workload?" "Should we start with 9B, 35B, or 397B?" "How does it pair with OpenClaw?" — 1–2 weeks, written report deliverable.

👉 Contact us — AI consulting (evaluation)

Stand up Ornith-1.0 with PoC build, fine-tuning, inference-engine setup, and quantization in 4–8 weeks. GPU sizing (RTX 5090 single-card vs B200 node) and ROI measurement included.

👉 Contact us — PoC / production SI

Ongoing support for model-update tracking, quant re-tuning, new-release evaluation, KPI monitoring, and OpenClaw backend swaps for Ornith-1.0.

- For OpenClaw-deployed sites: OpenClaw maintenance — Light ¥9,800 / Standard ¥19,800 / Premium ¥49,800 per month - AI-consulting continuous support: Light ¥30,000/mo (monthly meeting + new-model tracking) / Standard ¥80,000/mo (bi-weekly + prompt tuning + monthly KPI review + training & FAQ updates) / Premium on request

👉 Contact us — OpenClaw maintenance

Q1. Is Ornith-1.0 OK for commercial use? A. Yes — MIT. Fully commercial, modifiable, redistributable, no regional restrictions, no extra contract. Re-check the upstream Qwen 3.5 / Gemma 4 licenses for any inherited obligations.

Q2. Difference vs Chinese OSS (Kimi K2.7-Code / GLM-5.2)? A. The data-sovereignty story is different. Ornith-1.0 is MIT, no regional restrictions, and there's no Chinese-routed-API concern. SWE-Bench Verified is in the same band as Kimi K2.7-Code (82.4% on 397B), with a clearer US (presumed) + MIT procurement story. See our Kimi K2.7-Code coverage.

Q3. Can a single RTX 5090 run it? A. The 9B runs comfortably (~19GB at BF16 — half of the 5090's 32GB). 35B runs on a single card at GGUF Q4 / Q5, though full 262K context wants multiple cards. The 397B cannot run on a single card — you need an 8×80GB GPU node or 4×B200.

Q4. Does SWE-Bench Verified 82.4% really beat Claude Opus 4.8? A. It's DeepReinforce's self-reported number — not yet on the public SWE-Bench leaderboard, no independent validation. Mandatory PoC measurement on your own code before production. That said, claiming 69.4% even on the 9B is noteworthy.

Q5. How does it integrate with OpenClaw? A. Officially compatible — ClawEval is in DeepReinforce's published benchmark suite (77.1% on 397B). Existing customers on our OpenClaw setup service can swap Ornith-1.0 in as the backend LLM. Our [OpenClaw maintenance plans](../services/openclaw-setup) cover this migration work.

Q6. Which size — 9B, 35B, or 397B? A. 9B for individual / SMB PoC, 35B for mid-market production, 397B for enterprise flagship. 9B runs on one GPU so PoC cost is minimal; 35B is the price-performance sweet spot; 397B is SOTA but wants 8×H100-class.

Q7. What does "RL-optimized scaffolding" actually mean in practice? A. Imagine the model has internalized agent-runtime experience in its weights. For the same prompt, Ornith-1.0 has already learned how to decompose problems and when to call which tools — so in an agent framework, it should reach the right answer in fewer iterations than a bare LLM. The numerical evidence is currently DeepReinforce's own published claim; third-party comparative data is still thin.

Q8. Relationship to CUDA-L1 / CUDA-L2 / IterX? A. All products of the same RL-research organization, at different layers. CUDA-L1/L2 is GPU-kernel optimization; IterX is inference optimization; Ornith-1.0 is LLM self-improvement itself. Reading Ornith as DeepReinforce's RL playbook applied to the LLM layer is the clean mental model.

Ornith-1.0 is DeepReinforce's MIT open-weights LLM family specialized for agentic coding, released June 26, 2026. Three sizes (9B / 35B / 397B), all at 262K context, RL-optimized scaffolding, SWE-Bench Verified 82.4% on the 397B (vendor-reported).

Why it matters: 1. MIT + no regional restrictions — fully sidesteps Chinese-OSS cross-border-data scrutiny 2. Complete three-size lineup — individual workstation to enterprise flagship in one family 3. Published on the standard SWE-Bench suite — easier to compare than Kimi K2.7-Code's vendor-only numbers 4. ClawEval treated as first-class — strong natural fit for our OpenClaw service customers

Caveat: vendor-reported benchmarks; independent validation still to come. PoC measurement is non-negotiable before production.

At Oflight, we cover Ornith-1.0 (and other local LLMs) end-to-end — evaluation, PoC, OpenClaw integration, and ongoing maintenance. Use the three inquiry funnels above to get in touch.

Primary: - DeepReinforce official Ornith-1.0 page - Hugging Face collection Ornith-1.0 - Hugging Face: Ornith-1.0-9B - Hugging Face: Ornith-1.0-35B - Hugging Face: Ornith-1.0-397B - DeepReinforce GitHub - DeepReinforce X (@deep_reinforce) DeepReinforce prior work: - CUDA-L1 GitHub - CUDA-L2 GitHub - CUDA-L1 official blog - MarkTechPost — CUDA-L1 coverage Benchmarks: - SWE-Bench official Related Oflight columns: - Kimi K2.7-Code (same-window competitor) - Local LLM June 2026 Update - Loop Engineering — scaffolding origin - Sakana Fugu — different orchestration approach - Cognition FrontierCode benchmark - PLaMo 3.0 Prime - Claude Code Agent View - Cursor Automations - cmux (Manaflow) - Liquid AI Japanese-specialized models Oflight services: - AI Consulting - OpenClaw setup - Software Development Inquiries: - AI consulting (evaluation / PoC) - OpenClaw + Ornith integration / maintenance - Custom software development / SI Note: DeepReinforce's HQ location, executive team, funding status, hosted SaaS API availability, Japan-region endpoint, planned SWE-Bench leaderboard registration, third-party scores on Cognition FrontierCode / Aider polyglot / LiveCodeBench, official VS Code / JetBrains extensions, and general-knowledge benchmark scores (AIME / GPQA-Diamond / MMLU) were not confirmable as of June 26, 2026. Verify with the Hugging Face model cards and DeepReinforce GitHub before production decisions.