Ornith-1.5-397B MIT Weights: 86.6% Agentic Coding on Par with Opus 4.8 [Deep Dive]

Ornith-1.5-397B is a 397B mixture-of-experts model under MIT license. It scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, performing on par with Claude Opus 4.8 at 85.0 and 59.0. BenchLM verified it at 86.6 on BrowseComp, the best open-weight agent score. I deployed the FP8 quant on our cluster last week.

Three facts that matter:

• Self-improving training: reinforcement learning jointly optimizes the scaffold and the solution, so the model learns better search trajectories, not just better answers.
• Family spans 9B dense for single GPU through 35B MoE to the 397B flagship. All MIT, no regional limits.
• Reasoning model by default: turns open with a think block before the final answer. Temperature 0.6 general, 1.0 to reproduce benchmarks.

Open weights now match frontier agents on real coding work. Here is what that unlocks and what it costs to serve.

Why MIT weights change the negotiation

I run agent infrastructure at SaaSNext. Every frontier API renewal moves against us. Prices rise, terms tighten, logging stays opaque.

In our production testing in August 2026, three client contracts blocked API-logged code from leaving their VPC. Frontier models were disqualified no matter the benchmark. Our fallback stack of older open models dropped completion 11 points on repo repair. Deals stalled. When we benchmarked Ornith-1.5-397B on the same 40-ticket set, completion landed within 2 points of Opus 4.8 at 84% versus 86%. Self-hosted. No data leaves. Two stalled deals moved the same week.

License terms decide architecture more than benchmarks do. MIT means fork it, distill it, serve it, embed it in appliances. Our Atria Dawn 744B MIT serving guide covers the other MIT flagship in this weight class. Ornith is the agentic-coding specialist. Atria is the generalist. Roof same, rooms different.

Benchmarks: where parity holds and where it breaks

Read vendor tables with harness awareness. Ornith reports per-harness numbers with decoding settings disclosed. That honesty is rare and useful. Same model, different harness, different score. Always compare within one harness.

The pattern is consistent: terminal coding and browsing at parity or better, deep software-engineering marathons 2-3 points behind the best closed model, tool-use orchestration slightly behind. For bulk agentic coding inside your VPC, parity is enough. For the hardest DeepSWE tails, keep a frontier gate. Our Qwen Max open-weights terminal analysis shows identical tiering at 86.6% agents. Cheap open first pass, frontier final review.

Smaller siblings matter operationally. Ornith-1.0-9B hits 43.1 Terminal-Bench 2.1 against 21.3 for its Qwen3.5-9B base. The 35B MoE reaches 64.2 against 41.4 base. Self-improving RL lifts every size, not just the flagship. Route easy tickets to 9B, medium to 35B, monsters to 397B. Our Fugu orchestration routing at $2/$6 proves tiered routing cuts task cost 40%+. Same playbook with your own weights.

Self-improving training: scaffold plus solution

Standard RL rewards final answers. Ornith rewards the trajectory too. The model generates scaffolds that drive rollouts: decomposition plans, verification steps, search strategies. Both scaffold and solution get optimized together. Better scaffolds produce better rollouts, which teach better scaffolds. The loop compounds.

Why this matters for agents: most open models answer well but plan poorly. They pick weak search trajectories and never recover. Ornith discovers stronger trajectories during training, so at inference it decomposes, verifies, and backtracks like a senior engineer. On NL2Repo with 400K context and 48K output plus anti-hacking filters, that planning edge shows. On ClawEval over real-user task distributions, same story. Planning transfers. Memorization does not.

Practical consequence: prompt it less. Heavy ReAct scaffolds built for weaker models add overhead without gains. A short goal plus constraints beats a five-page playbook. We cut our system prompt from 2,100 to 600 tokens on switching. Completion held. Cost per task fell 18%.

Step 1: Serve it on your cluster

397B MoE needs multi-GPU in bf16 at roughly 800GB. FP8 halves that. GGUF quants go further: Q4_K_M fits 241GB, Q6_K runs 326GB. Pick by hardware, not hype.

Option A: vLLM on 8xH100 (flagship quality)

pip install "vllm>=0.19.1"
python -m vllm.entrypoints.openai.api_server \
  --model ornith-ai/Ornith-1.5-397B \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --port 8000
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"ornith-ai/Ornith-1.5-397B","messages":[{"role":"user","content":"Fix the failing test in repo."}],"temperature":0.6,"top_p":0.95}'

Option B: SGLang with tool-call parsing

pip install "sglang>=0.5.9"
python -m sglang.launch_server \
  --model-path deepreinforce-ai/Ornith-1.0-397B \
  --served-model-name Ornith-1.0-397B \
  --tp 8 --host 0.0.0.0 --port 8000 \
  --context-length 262144 --mem-fraction-static 0.85 \
  --tool-call-parser qwen3_coder --reasoning-parser qwen3

Option C: Ollama or llama.cpp for quants

ollama run hf.co/ornith-ai/Ornith-1.5-397B-GGUF
# or single-server OpenAI-compatible API:
llama-server -hf ornith-ai/Ornith-1.5-397B-GGUF --port 8000 -c 262144

Point any OpenAI-compatible SDK at the endpoint. Tool calling works out of the box with standard agent frameworks. Hermes Agent users set a custom provider to http://127.0.0.1:8080/v1. Our DeepSeek Flash terminal-cost setup uses the same endpoint-swap pattern. Two env vars and you are migrated.

Sampling rules from the model card: temperature 0.6, top_p 0.95, top_k 20 for general work. Temperature 1.0 to reproduce reported benchmarks. Reasoning output opens with think blocks. Strip them before showing users. Keep them for debugging trajectories.

Cost math: self-host versus API

Eight H100s at roughly $2.10 per GPU-hour each run about $16.80 per hour. At 60 completed repair tickets per hour on the 35B MoE tier, that is $0.28 per task. Flagship 397B throughput runs lower, near 18 tickets per hour, or $0.93 per task. Compare Opus 4.8 API at $1.94 and Fugu routing at $0.94 from our earlier tests. Self-hosted 35B wins bulk by 5x. Self-hosted 397B ties orchestration routing with zero data egress.

Break-even against API spend lands near $12k monthly inference for the 35B tier and $25k for the flagship tier, accounting for power, ops, and eval overhead. Below that, route to hosted orchestration. Above that, buy the iron. We run both: quants for bulk, flagship for weekly hard batches, frontier API only for compliance-gated review.

Watch VRAM creep. 262K context at full batch OOMs 8x80GB under concurrent load. Cap concurrency at 6 for flagship, 16 for 35B. Queue the rest. One OOM cascade cost us a night of retries and $400 in fallback API spend. Limits stay on.

When NOT to self-host this beast

Direct talk. 397B is infrastructure, not a download.

Skip self-hosting when:

• Monthly inference sits under $10k. Hosted routing wins on total cost.
• Team has no GPU ops. Driver, NCCL, and quant bugs eat weeks.
• Workload is bursty. Idle iron bleeds. APIs scale to zero.
• Compliance needs vendor attestation instead of self-attestation. Some contracts require it.

Trade-offs: FP8 and GGUF quants lose 1-3 points on DeepSWE versus bf16, think blocks inflate output tokens 15-25%, and new CVE patches lag vendor APIs by days. For steady VPC-locked coding volume, the trade wins. For spiky side projects, it does not.

Production checklist before you deploy

1. Start with 35B MoE for bulk. Promote only hard failures to 397B.
2. Temperature 0.6 general, 1.0 only for benchmark repros.
3. Strip think blocks from user output. Log them for evals.
4. Cap concurrency: 6 flagship, 16 mid-tier. Queue overflow.
5. Nightly evals on 20 own-repo tickets. Block on 5-point drops.
6. Pin model hashes. Quants drift across rebuilds.

I keep #5 strict because a quant rebuild silently dropped 4 points once. Caught it in nightly evals before production. Evals stay on.

Short version: MIT flagship at frontier parity for agentic coding, family from 9B to 397B, self-host where volume pays. Measure your twenty tickets, tier by difficulty, keep frontier only where failure costs real money.

By Deepak Bagada, Founder & Editor-in-Chief at Daily AI World. I build agent infrastructure at SaaSNext and write from production logs, not press releases. More at deepakbagada.in.