DeepReinforce 开源 Ornith-1.0 编程模型族，可自学习 RL 脚手架

DeepReinforce has released Ornith-1.0, an open-source model family built for agentic coding. The lineup spans four sizes, from a 9B dense model to a 397B mixture-of-experts flagship. Every checkpoint ships under the MIT license on Hugging Face. The models are post-trained on top of pretrained Gemma 4 and Qwen 3.5. Most coding agents pair a model with a fixed, human-designed harness. Ornith-1.0 instead learns to write its own. The DeepReinforce research team reports state-of-the-art results among open models of comparable size. TL;DR Ornith-1.0 ships in 9B, 31B, 35B-MoE, and 397B-MoE sizes under MIT, built on Gemma 4 and Qwen 3.5. The model learns its own scaffold during RL, jointly optimizing the harness and the solution. Ornith-1.0-397B tops Claude Opus 4.7 on both headline benchmarks, but not Opus 4.8 or the larger GLM-5.2-744B. Three layers — fixed trust boundary, deterministic monitor, frozen LLM judge — guard against reward hacking. What is Ornith-1.0? Ornith-1.0 is a set of reasoning models tuned for coding agents. The variants are 9B Dense, 31B Dense, 35B MoE, and 397B MoE. The 35B model is mixture-of-experts and activates roughly 3B parameters per token. FP8 and GGUF builds are also published for faster local serving. Each model is a reasoning model. Replies open with a <think> block before the final answer. The serving recipes enable a reasoning parser, so that trace returns in a separate reasoning_content field. The models also emit well-formed tool calls for agent loops. Deployment is straightforward. The 9B model is about 19GB in bf16 and serves on a single 80GB GPU. Serving recipes target vLLM, SGLang, and Transformers. Each model exposes an OpenAI-compatible endpoint. Standard agent frameworks therefore work without code changes. Interactive Explainer </button> <button class="btn gho" id="resetBtn">Reset</button> </div> <div class="stepout" id="stepOut">Step 0 — untrained policy with a fixed, hand-written harness.</div> </div> <!-- PANEL 2: BENCH --> <div class="panel" data-panel="bench"> <div class="lead">Vendor-reported scores from DeepReinforce. Pick a model tier and a benchmark. Ornith is highlighted in green. Higher is better.</div> <div class="seg"><span class="lab">Model tier</span> <div class="chip on" data-tier="t397">397B flagship</div> <div class="chip" data-tier="t35">35B MoE</div> <div class="chip" data-tier="t9">9B dense</div> </div> <div class="seg" id="benchChips"><span class="lab">Benchmark</span></div> <div class="chart" id="chart"></div> <div class="foot-note" id="benchNote"></div> </div> <!-- PANEL 3: DEFENSES --> <div class="panel" data-panel="def"> <div class="lead">A model that writes its own scaffold could cheat the verifier. DeepReinforce describes three defense layers. Tap each to expand.</div> <div class="layers"> <div class="layer open"><div class="lh"><span class="num">1</span><span class="lt">Fixed trust boundary</span><span class="more">tap</span></div><div class="lb">The environment, tool surface, and test isolation are immutable and outside the model's reach. The model evolves only its inner policy scaffold — memory, error-handling, and orchestration logic.</div></div> <div class="layer"><div class="lh"><span class="num">2</span><span class="lt">Deterministic monitor</span><span class="more">tap</span></div><div class="lb">A rule-based monitor flags any attempt to read withheld paths, modify verification scripts, or invoke unsanctioned tools. Such trajectories get zero reward and are excluded from the advantage computation.</div></div> <div class="layer"><div class="lh"><span class="num">3</span><span class="lt">Frozen LLM judge</span><span class="more">tap</span></div><div class="lb">Because intent-level gaming can happen inside the allowed tool surface, a frozen LLM judge acts as a veto on top of the verifier — not as the primary reward signal.</div></div> </div> </div> <div class="ftr"><span>Source: <a href="https://deep-reinforce.com/ornith_1_0.html" target="_blank" rel="noopener">deep-reinforce.com</a> · MIT licensed · numbers vendor-reported</span><span><b>Marktechpost</b> · AI Dev Signals</span></div> <script> (function(){ var root=document.getElementById('mtp-ornith-demo'); /* tabs */ root.querySelectorAll('.tab').forEach(function(t){ t.addEventListener('click',function(){ root.querySelectorAll('.tab').forEach(function(x){x.classList.remove('on')}); root.querySelectorAll('.panel').forEach(function(x){x.classList.remove('on')}); t.classList.add('on'); root.querySelector('.panel[data-panel="'+t.dataset.p+'"]').classList.add('on'); resize(); }); }); /* loop sim */ var step=0,reward=0.08,timer=null; var scaffs=[ 'Baseline harness: linear retries, no memory.', 'Adds scratchpad memory across tool calls.', 'Adds error-triage branch before re-edit.', 'Reorders: read tests, then plan, then patch.', 'Caches sub-results; prunes dead branches.', 'Task-specific orchestration emerges automatically.']; var outs=[ 'Fixed harness, no learning yet.', 'Fewer redundant file reads observed.', 'Recovers from failed edits more often.', 'Higher first-pass test success.', 'Shorter trajectories, same accuracy.', 'Stable high-reward scaffold selected.']; var nodes=root.querySelectorAll('.node'); function lightSeq(cb){ var i=0;nodes.forEach(function(n){n.classList.remove('act')}); var iv=setInterval(function(){ nodes.forEach(function(n){n.classList.remove('act')}); nodes[i].classList.add('act');i++; if(i>=nodes.length){clearInterval(iv);setTimeout(function(){nodes.forEach(function(n){n.classList.remove('act')});cb&&cb();},260);} },220); } function doStep(){ if(step>=5){return;} step++; lightSeq(function(){ reward=[0.08,0.27,0.43,0.58,0.69,0.77][step]; root.querySelector('#rFill').style.width=(reward*100)+'%'; root.querySelector('#rVal').textContent=reward.toFixed(2); root.querySelector('#scaffTxt').textContent=scaffs[step]; root.querySelector('#outTxt').textContent=outs[step]; root.querySelector('#stepOut').innerHTML='Step '+step+' — <b>scaffold mutated</b>; reward propagated to both stages.'; resize(); }); } root.querySelector('#stepBtn').addEventListener('click',doStep); root.querySelector('#autoBtn').addEventListener('click',function(){ if(timer){clearInterval(timer);timer=null;this.textContent='Auto-run <img src="https://s.w.org/images/core/emoji/17.0.2/72x72/25b6.png" alt="▶" class="wp-smiley" style="height: 1em; max-height: 1em;" />';return;} this.textContent='Pause <img src="https://s.w.org/images/core/emoji/17.0.2/72x72/23f8.png" alt="⏸" class="wp-smiley" style="height: 1em; max-height: 1em;" />';var b=this; timer=setInterval(function(){if(step>=5){clearInterval(timer);timer=null;b.textContent='Auto-run <img src="https://s.w.org/images/core/emoji/17.0.2/72x72/25b6.png" alt="▶" class="wp-smiley" style="height: 1em; max-height: 1em;" />';}else{doStep();}},1400); }); root.querySelector('#resetBtn').addEventListener('click',function(){ if(timer){clearInterval(timer);timer=null;root.querySelector('#autoBtn').textCon