What is Logos? The Fine-Tuning Service Where You Keep the Weights

What is Logos?

Logos is a fine-tuning-as-a-service offering from Daedalus Development Group. You pick a base model from a curated library of 45 open-weight LLMs across 16 families, describe your objective, upload your training data, and pay a one-time flat fee. We train, evaluate, and deliver the actual weight files to your inbox.

The defining characteristic: you own the weights when we're done. We do not host the model. We do not provide an inference endpoint. You run the trained model wherever you want — locally, on your own cloud, or through any inference provider (vLLM, llama.cpp, Ollama, Together, Fireworks, RunPod). There is no vendor lock-in by design.

Why that matters

Most commercial fine-tuning services are gated inference businesses. OpenAI's fine-tuning API produces a private model that only runs through OpenAI's endpoints at per-token rates. Together AI, Replicate, and Fireworks fine-tune open models but monetize the hosted inference. Your fine-tuned model is effectively rented, not owned — and if the hosting provider raises prices, deprecates the underlying base, or cuts off access, your work is stranded.

Logos flips that model. The training is the product. The weights are the deliverable. What you do with them is your business.

What you get when we're done

Every Logos engagement ships:

Merged full-precision weights in .safetensors format — the entire model, ready to load into any modern inference framework
LoRA adapter (for LoRA / QLoRA trainings) — smaller file, deploy on top of the base model, swap in and out
Evaluation report as HTML — before/after metrics on your tasks, sample outputs, recommendation memo
Training configuration as YAML — reproducible settings plus training logs
Inference snippets — copy-paste Python, vLLM, and llama.cpp examples so you can serve the model the same day the link arrives
Atelier tier only: direct engineer consultation via email during delivery

Weights are delivered via a time-limited signed URL — 14 days on Starter and Studio, 30 days on Atelier. After you download, the model is yours forever. We delete our copy when the window closes.

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	SMB production work, most commercial use cases	≤ 35B dense / 50B MoE	100,000	3
Atelier	$3,499	Production deployments, 70B class and up, frontier MoE	Up to 1.04T MoE (+surcharge)	Unlimited	Custom
Custom	Quoted	Off-catalog models, full-parameter FT, multi-job programs, NDA engagements	Any	Any	Any

Prices are one-time fees, not subscriptions. Larger and reasoning-heavy models (Qwen3 235B, Nemotron Ultra 253B, GLM-5.1, Kimi K2.5) carry a transparent per-model surcharge shown on the model selector — no opaque upsells.

How Logos differs from OpenAI, Together, and Replicate

At a glance:

	Logos	OpenAI Fine-Tuning	Together AI	Replicate
Do you own the weights?	Yes	No	Yes (export)	Yes (export)
Hosted inference required?	No	Yes	Optional	Default
Model catalog	30+ open-weight	GPT family only	~15 open	Many
Pricing model	Flat one-time fee	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 inference endpoint on day one with zero ops work, Together or Fireworks are better fits. If you want the weights, Logos ships them. A longer post comparing these services in depth is coming next in this silo.

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 — trains every weight of the base model. Maximum quality ceiling, highest compute cost. Available via the Custom tier (including full-FT on Mixtral 8x22B, Qwen 235B, and other off-catalog configurations). Atelier-tier catalog engagements run LoRA/QLoRA; full-FT on a specific Atelier model is quoted as Custom work.

For most production use cases, a well-designed LoRA or QLoRA run on the right base model beats a full-parameter run on the wrong one. The Atelier tier's engineer consultation is designed specifically to help you pick correctly — it saves more than it costs.

The model library, at a glance

Forty-five open-weight models across 16 families. Headline coverage:

Meta — Llama 3.1 8B, Llama 3.3 70B
Alibaba (Qwen 2.5) — 3B / 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
Mistral AI — Mistral 7B v0.3, Nemo 12B, Small 24B, Mixtral 8x7B, Mixtral 8x22B
Google — Gemma 3 12B / 27B, Gemma 4 4B / 26B-MoE / 31B (multimodal)
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, #1 on SWE-Bench Pro)
Moonshot AI — Kimi K2.5 (1.04T MoE, vision + agent swarm)
Nous Research (Hermes) — Hermes 3 Llama 3.1 8B, Hermes 4.3 36B (ByteDance Seed base), Hermes 3 Llama 3.3 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 available on top of any tier:

Unsloth Acceleration ($249) — trains the same-quality model 2–3× faster via memory-efficient kernels. Pays for itself on turnaround-sensitive projects.
Function Calling / Tool Use ($499) — synthesizes tool-call training pairs from your schema, validates schema conformance on 200 generated calls. Eliminates the single most common failure mode for agentic deployments.
Quantization Package ($349) — delivers three precision variants (FP16, AWQ INT8, GGUF Q4_K_M) so the model runs everywhere from datacenter inference to edge devices.
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 your training data contains protected health information, financial records, government-regulated material, or anything that cannot leave your environment, Logos is not the right service. For HIPAA / SOC 2 / PCI / ITAR / air-gapped engagements, use Erkos — DDG's on-prem secure fine-tuning service, built on the IronClaw security framework. Erkos runs in your facility, your data never touches a shared network, and weights stay behind your firewall.

Logos is optimized for prototyping, commercial non-regulated production, and open-weight model iteration speed. Erkos is optimized for everything else.

Frequently asked questions

What is the cheapest Logos tier?

Starter at $349. It covers LoRA fine-tuning on any model up to 14B parameters, up to 10,000 training examples, and one training epoch. Best fit for prototypes, style adaptation, and niche domain experiments.

Do I actually keep the fine-tuned weights?

Yes. You receive a merged full-precision .safetensors file plus the LoRA adapter via a signed download URL (14 days on Starter and Studio, 30 days on Atelier). After you download, the model is yours to run anywhere. We delete our copy when the window closes.

Can I run the fine-tuned model on OpenAI or Together AI after Logos delivers it?

You can't run it on OpenAI (different weight format, not their model). You can run it on Together, Fireworks, RunPod, vLLM, llama.cpp, Ollama, MLX, or any inference provider that accepts .safetensors or GGUF — we deliver inference snippets for the common paths.

How does Logos compare to OpenAI fine-tuning?

OpenAI fine-tunes GPT-family models you can only access through OpenAI's API at per-token rates. Logos fine-tunes open-weight models (Llama, Qwen, Mistral, Gemma, Phi, DeepSeek, Nemotron, GLM, Kimi) and ships you the weights. Flat-fee pricing vs. per-token metering. Full comparison here →

Is my training data kept private?

Training data is uploaded to isolated per-job storage, used only for your job, and deleted on job completion. We do not reuse customer data for any other model or purpose. For HIPAA / SOC 2 / PCI / ITAR / air-gapped environments, use Erkos — DDG's on-prem secure fine-tuning service.

What's the turnaround time?

Dependent on dataset size, base model, and current GPU capacity. A 10,000-example LoRA on a 7B model is typically hours. A full-parameter 70B with DPO on 100K examples is days. We estimate specifically at intake when we've seen your data — not before.

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?

Yes — the delivered weights run with any agent framework (LangChain, CrewAI, opentine, etc.). For teams building on top of opentine (DDG's fork-able agent runtime), Logos-trained local weights compose naturally inside run trees.

Next steps

Browse the Logos pricing and model selector
Pay with PayPal, card (Stripe + Link), or crypto (NOWPayments: BTC / ETH / USDC / XMR / ZEC and more)
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 — the next is a detailed comparison of Logos against OpenAI fine-tuning, Together AI, and Replicate. After that, a model-by-model ranked guide to all 30+ base models in the catalog.