George Sarris (0xcircuitbreaker) · · 9 min read
What is Logos? The Fine-Tuning Service Where You Keep the Weights
Logos scopes open-weight fine-tuning and delivers agreed artifacts. Price, hardware feasibility, licenses, and runtime compatibility are confirmed at intake.
In this article
What is Logos?
Logos is a fine-tuning-as-a-service offering from Daedalus Development Group. The catalog currently lists 45 open-weight model entries across 17 families. Intake verifies model access, license, training method, data readiness, capacity, and target runtime before DDG accepts a fixed scope. Approved jobs deliver downloadable model artifacts through the agreed delivery channel.
The defining characteristic: you receive the agreed weights or adapter. We do not host an inference endpoint. Your rights remain subject to the base-model license, dataset rights, and engagement terms. Deployment depends on architecture, format, precision, runtime support, and each provider's upload policy; .safetensors may need conversion or packaging for llama.cpp, Ollama, MLX, or a hosted provider.
Why that matters
Fine-tuning services differ in artifact portability. OpenAI's current pricing page says its fine-tuning platform is winding down and unavailable to new users; existing fine-tuning users may create training jobs for the coming months, and resulting models remain tied to OpenAI's service. Together AI explicitly supports merged and adapter checkpoint downloads, including Hugging Face Hub export. Replicate export depends on the trainable model and workflow. It is inaccurate to describe all of these open-model artifacts as “rented.”
Logos is organized around a scoped training engagement and an artifact delivery rather than a required hosted-inference subscription. How the artifact may be used still depends on the base-model license, dataset rights, and accepted engagement terms.
What you get when we're done
Deliverables vary by tier and accepted scope:
- Starter lists merged full-precision weights, a PDF baseline-comparison report, inference examples, and a remote deployment session. A LoRA adapter is included when that is the accepted method.
- Studio includes the Starter deliverables plus LoRA or QLoRA options, quantized and adapter artifacts where compatible, and a reproducible training-configuration export.
- Atelier adds a use-case-specific HTML evaluation harness/report, an included revision round, and direct engineer consultation.
- Custom deliverables are negotiated for the engagement rather than implied by a standard checklist.
Standard delivery windows are listed as 14 days for Starter and Studio and 30 days for Atelier. The repository does not establish an automated retention-enforcement control, so the accepted order must state the actual retention and deletion obligation. Continued possession and use of a downloaded artifact remain subject to the applicable licenses and terms.
Who is Logos for?
Logos is built for the team that has answered two questions:
- Does a specific open-weight model fit my inference budget? (If yes: Llama 3.3 70B, Qwen 2.5 72B, Mixtral 8x22B, and smaller options are all in the catalog.)
- Do I have training data that makes the base model better at my task? (Domain Q&A, style adaptation, classification, function calling, RAG optimization, code generation — all common patterns.)
If you're still exploring which base model to use, the Atelier tier includes engineer consultation. If you're prototyping, Starter at $349 gets you a LoRA on a small model in the first week.
The four tiers
| Tier | Price | Ideal for | Max model size | Max examples | Epochs |
|---|---|---|---|---|---|
| Starter | $349 | Prototypes, side projects, style adaptation | ≤ 14B dense | 10,000 | 1 |
| Studio | $1,499 | SMBs, agencies, and product teams | ≤ 35B dense / 50B MoE | 100,000 | 3 |
| Atelier | From $3,499 | Larger-model projects accepted after capacity review | Model-specific (+surcharge) | Scoped at intake | Custom |
| Custom | Quoted | Off-catalog models, full-parameter FT, multi-job programs, NDA engagements | Feasibility-reviewed | Scoped | Scoped |
These were the listed intake prices on August 24, 2026, not offers to train every catalog model at the base tier price. Larger models carry model-specific surcharges and may require a Custom quote after a memory plan and pilot. Public checkout is paused; each job starts with intake review for model access, data readiness, capacity, license, and delivery scope.
How Logos differs from OpenAI, Together, and Replicate
At a glance:
| Logos | OpenAI Fine-Tuning | Together AI | Replicate | |
|---|---|---|---|---|
| Downloadable artifacts | Yes, as scoped | No | Yes (checkpoint export) | Model/workflow-specific |
| Hosted inference required? | No | Yes | Optional | Default |
| Model catalog | 45 listed entries | Platform winding down for new users | Changes over time | Trainable-model-specific |
| Pricing model | Scoped tier fee, plus disclosed model/add-on charges | Per-token training + per-token inference | Per-token training + optional hosted inference | Per-compute-second |
| Evaluation report | Included | Not included | Not included | Not included |
| Starting price | $349 | Usage-based | Usage-based | Usage-based |
The trade-off is real: if you want a hosted endpoint with managed operations, a hosting platform may be a better fit. If you want a scoped offline deliverable, Logos is designed around that outcome. Verify current provider features and prices before buying.
Training techniques — and when each makes sense
Logos supports three training techniques across the catalog:
- LoRA (Low-Rank Adaptation) — the fastest, cheapest, most reusable option. Produces a small adapter file that patches the base model. Works on every Logos tier. Default choice for Starter.
- QLoRA — quantized LoRA, drops memory footprint further. Enables training larger base models on smaller GPU budgets. Studio-tier default for models above 24B.
- Full-parameter fine-tuning — updates every trainable weight and has much larger compute and memory requirements. It is Custom-only and offered only after feasibility review; catalog inclusion does not guarantee that a frontier model can be full-fine-tuned on available capacity.
LoRA, QLoRA, and full-parameter training have different quality, stability, cost, and deployment trade-offs. The right choice should be made from a baseline evaluation and a measured pilot rather than a universal quality claim.
The model library, at a glance
Forty-five catalog entries across 17 open-weight model families. Public checkout is paused, so these are intake candidates rather than an assurance that every listed job can be accepted. Headline coverage:
- Meta — Llama 3.1 8B, Llama 3.3 70B
- Alibaba (Qwen 2.5) — 7B / 14B / 32B / 72B
- Alibaba (Qwen 3 dense + MoE) — 8B / 14B / 30B-A3B MoE / 32B / 235B-A22B MoE
- Alibaba (Qwen 3.5) — 2B / 4B / 9B / 27B / 35B-A3B MoE / 122B-A10B MoE / 397B-A17B MoE
- Alibaba (Qwen 3.6) — 27B / 35B-A3B MoE
- Qwen / Alibaba Cloud (Qwen 3.8) — official 27B native vision-language model
- Mistral AI — Mistral 7B v0.3, Nemo 12B, Small 24B, Mixtral 8x7B, Mixtral 8x22B
- Google — Gemma 3 12B / 27B; Gemma 4 E4B (vision + audio), 26B-A4B / 31B (vision, not audio)
- Microsoft — Phi-4 14B
- DeepSeek — R1-Distill-Llama-8B, R1-Distill-Qwen-32B
- NVIDIA — Llama-Nemotron Nano 8B / Super 49B / Ultra 253B
- Z.ai (Zhipu) — GLM-5 (744B), GLM-5.1 (754B; vendor reports 58.4 on SWE-Bench Pro under its harness)
- Moonshot AI — Kimi K2.5 (1.04T MoE, vision + agent swarm)
- Nous Research (Hermes) — Hermes 3 Llama 3.1 8B, Hermes 4.3 36B (Apache-2.0), Hermes 3 Llama 3.1 70B
- MiniMax — MiniMax-M2 (230B-A10B MoE, agentic/coding), MiniMax-M1 (456B-A45.9B MoE, 1M context)
Every model lists its minimum tier, surcharge (if any), and strength on the Logos selector. No hidden pricing.
Add-ons that actually earn their keep
Five add-ons are listed, subject to the selected tier, model, data, and accepted scope:
- Unsloth Acceleration ($249) — uses memory-efficient kernels where compatible. Speed and output equivalence depend on model, hardware, sequence length, trainer version, and settings; results are reported from the actual run.
- Function Calling / Tool Use ($499) — synthesizes tool-call training pairs from your schema and tests schema conformance on 200 generated calls. This evaluates one important failure class; it does not establish end-to-end agent reliability.
- Quantization Package ($349) — delivers compatible precision variants selected at intake, such as FP16/BF16, AWQ 4-bit, or GGUF Q4_K_M. Not every architecture supports every format or runtime.
- Custom Evaluation Suite ($699) — we design 50–200 task-specific eval prompts, run before/after baselines, deliver a detailed comparison report. Worth it when "is this actually better?" needs a rigorous answer.
- DPO Preference Tuning ($899) — Direct Preference Optimization on customer-supplied preference pairs. Runs on top of SFT to align tone, style, or refusal calibration when more data alone won't fix it. Studio and Atelier only.
When to use Erkos instead
If data cannot leave your environment, request a separately scoped on-prem engagement. HIPAA, PCI DSS, ITAR, and SOC 2 are different legal, contractual, and assurance regimes; SOC 2 is an attestation, not a category of regulated data. On-prem deployment or use of IronClaw does not by itself establish compliance or certification. Require a written scope covering controls, responsibilities, data flows, subprocessors, evidence, and legal review.
Logos is designed for accepted projects that fit its published data-handling scope. Anything involving regulated or highly sensitive data requires case-specific security and legal review; no service name is a blanket compliance guarantee.
Frequently asked questions
What is the cheapest Logos tier?
Starter is listed at $349. It covers LoRA fine-tuning on an eligible catalog model up to 14B parameters, up to 10,000 training examples, and one training epoch, subject to intake acceptance. It is aimed at prototypes, style adaptation, and niche domain experiments.
Do I actually keep the fine-tuned weights?
You receive the artifact specified in the accepted scope: merged weights, an adapter, or both where technically and legally permitted. Retention and deletion dates must appear in the order terms. Deployment rights and compatibility remain subject to the base-model license, dataset rights, format, architecture, and runtime.
Can I run the fine-tuned model on OpenAI or Together AI after Logos delivers it?
You cannot upload arbitrary open weights to OpenAI's hosted GPT service. Other providers accept only supported architectures and formats. vLLM may load compatible .safetensors; llama.cpp and many Ollama paths require GGUF conversion; MLX requires compatible conversion. Provider upload policy is confirmed per delivery.
How does Logos compare to OpenAI fine-tuning?
OpenAI's current pricing page says its fine-tuning platform is winding down and unavailable to new users. Existing fine-tuning users may create jobs for the coming months, while fine-tuned models remain hosted until their base models are deprecated. Logos instead scopes open-weight work and delivers the agreed artifacts. Full dated comparison here →
Is my training data kept private?
The accepted order must state storage location, access controls, subprocessors, retention, deletion timing, backup handling, and whether data is reused. Do not rely on this blog as a privacy contract. Regulated or air-gapped work requires a separately documented security and legal scope.
What's the turnaround time?
Turnaround depends on token count, model, sequence length, method, evaluation plan, runtime, and current capacity. DDG provides an estimate only after intake and, for large jobs, a measured pilot. No universal “hours” or “days” promise applies.
What if the first run doesn't hit my target?
Atelier tier includes one revision round to tune hyperparameters. Starter and Studio are one-shot runs — we refund in full only if a job fails due to our infrastructure; once successful compute is consumed, fees are not refundable.
Can I fine-tune with my agent framework?
Potentially, after compatibility validation. An agent framework does not load every weight artifact automatically: its configured inference runtime or provider must support the exact architecture, tokenizer and chat template, precision or quantization, and artifact format. Conversion or a separately served endpoint may be required. OpenClaw and Hermes Agent are runtimes, while opentine records and analyzes agent runs; none of those names guarantees that a delivered Logos artifact will load directly. If you only need public skills or framework discovery, use the Hermes skills documentation. If the workflow needs customer channels, approvals, reports, or managed operations, route it through Icarus.
Next steps
- Browse the Logos pricing and model selector
- Request intake review before payment, dataset upload, or checkout
- Upload your data via signed link after order
- Receive weights, evaluation report, and inference snippets via signed download URL
More Logos-silo posts are already in the pipeline, including buyer guides for weight ownership, model selection, and when to pair Logos with Icarus or framework setup.