Self-Hosted LLM: DeepSeek and Qwen 2026

A model name is not a hardware plan. DeepSeek-R1 and Qwen checkpoints differ by parameter count, artifact size, context configuration, base model, and license. Ollama simplifies local serving, but reliable selection still requires an explicit tag and a benchmark on the hardware, quantization, context, and concurrency you will use.

TL;DR verdict

Use explicit Ollama tags. Current deepseek-r1:latest is an updated 8B distilled model; explicit 7B, 14B, 32B, 70B and 671B tags have distinct artifact sizes. Treat those sizes and model-card context values as screening inputs, not VRAM, speed, latency, or quality guarantees. License each checkpoint separately, then run a task-specific evaluation on the target machine before choosing a model for production work.

Key takeaways

The current deepseek-r1:latest tag points to an updated 8B distilled model; pin an explicit tag instead of inheriting future changes.
DeepSeek-R1 is MIT-licensed, while Qwen-derived distills retain Apache-2.0 base-model licensing qualifications.
Qwen2.5-Coder-32B-Instruct is a 32.5B-parameter Apache-2.0 checkpoint with a publisher-claimed 131,072-token full context.
Ollama supports parts of the OpenAI API, but unsupported fields and stateful Responses behavior require compatibility testing.
Artifact and context metadata are screening inputs. Benchmark on target hardware before making fit or capacity claims.
Publisher evaluations can inform a test plan; they do not establish one model as the universal choice for coding or reasoning.

At-a-glance table

Candidate	Current source-backed facts	What to verify locally
DeepSeek-R1 explicit tags	Ollama lists 7B, 14B, 32B, 70B, and 671B tags with different artifact sizes	Memory fit, prompt format, output quality, latency, and concurrency
`deepseek-r1:latest`	Current tag is an updated 8B distilled model	Whether an unpinned alias is acceptable for your deployment
DeepSeek-R1 checkpoints	Main model card uses MIT; distills retain base-model license qualifications	The checkpoint-specific license and notice files
Qwen2.5-Coder-32B-Instruct	32.5B parameters, Apache-2.0, publisher-claimed 131,072-token full context	Effective context, memory, coding tasks, and tool integration
Qwen3 in Ollama	Dense and mixture-of-experts variants with thinking-capable models	Exact tag, current controls, prompt behavior, and task fit
Ollama compatibility surface	Chat completions, streaming, tools, and reasoning controls are documented	Every request field, response field, error path, and state assumption

Choose the checkpoint, not just the family

Use explicit Ollama tags. Current deepseek-r1:latest is an updated 8B distilled model; explicit 7B, 14B, 32B, 70B and 671B tags have distinct artifact sizes.

An explicit tag prevents a deployment from silently changing when an alias moves. It also keeps benchmark results tied to the artifact that produced them.

ollama pull deepseek-r1:8b
ollama run deepseek-r1:8b

ollama pull deepseek-r1:14b
ollama run deepseek-r1:14b

Before pulling a larger tag, check its current Ollama library entry and available disk space. Artifact size alone does not equal total runtime memory.

Checkpoint-specific license review

License each checkpoint separately: DeepSeek-R1 is MIT, Qwen-derived distills retain Apache-2.0 base licensing, and Qwen2.5-Coder-32B-Instruct is Apache-2.0.

Keep the model card, repository notices, and any base-model qualifications with the deployed artifact. A shared family name does not establish the same obligations for another checkpoint.

License box

Checkpoint or family	Approved evidence	Review action
DeepSeek-R1	Main model card identifies MIT	Verify the exact checkpoint and notices distributed with it
Qwen-derived DeepSeek distills	DeepSeek card retains Apache-2.0 base-model qualifications	Review both DeepSeek and named base-model terms
Qwen2.5-Coder-32B-Instruct	Model card labels the checkpoint Apache-2.0	Preserve license and model-card attribution
Other Qwen3 tags	Ollama confirms distribution and thinking capability	Reopen the exact model card before deployment

Qwen2.5-Coder-32B-Instruct

Qwen2.5-Coder-32B-Instruct is a 32.5B-parameter Apache-2.0 model with a publisher-claimed 131,072-token full context; do not convert publisher benchmark claims into “best” recommendations.

The model card can shape a coding evaluation, but the workload should decide whether the checkpoint fits. Test repository-scale context, code generation, refactoring, explanation, tool calls, and the languages you maintain. Pin the serving configuration and record prompt templates so others can reproduce the results.

ollama pull qwen2.5-coder:32b
ollama run qwen2.5-coder:32b

Confirm the current Ollama tag and model provenance before relying on this command in automation.

Qwen3 and thinking-capable models

The current Ollama Qwen3 library documents dense and mixture-of-experts variants and thinking-capable models. It does not, in this evidence packet, establish the exact interactive control syntax reproduced by the older guide.

The legacy examples were:

>>> /think Write a recursive function to flatten nested lists in Python.
>>> /no_think What does the zip() function do?

Treat those lines as historical copy, not current instructions. Reopen the current Qwen3 and Ollama documentation before configuring a "thinking mode" or a "deep reasoning," workflow.

Hosted vs self-hosted

Area	Hosted model API	Self-hosted Ollama endpoint
Model selection	Provider-controlled catalog and aliases	Operator selects and pins local artifacts
Data path	Requests cross the provider boundary under its terms	Prompts can remain on operator-controlled infrastructure
Capacity	Provider plan and service limits	Hardware, quantization, context, concurrency, and queueing determine capacity
Cost	Current provider pricing and usage	Hardware, power, storage, administration, and utilization
Updates	Provider changes models and service behavior	Operator chooses when to pull and roll out a new artifact
Security	Shared with provider	Host, network, application, access control, logs, and model supply chain are yours

Ollama's compatibility boundary

Ollama supports parts of the OpenAI API, including chat completions, streaming, tools and reasoning control; it does not implement every field or stateful Responses behavior.

Partial compatibility does not mean every existing client can point to Ollama unchanged. Inventory the fields and endpoints your application uses, then test unsupported fields, streaming, tool calls, reasoning options, errors, and stateless versus stateful assumptions.

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-r1:8b",
    "messages": [{"role": "user", "content": "Summarize this incident report."}]
  }'

If an application points to http://localhost:11434, bind and expose the service according to your threat model. A local endpoint is not automatically private when the host, container network, proxy, or logs are misconfigured.

Maintenance burden

A useful local model service needs more than a successful response. Own the artifact inventory, disk growth, model provenance, license records, host updates, access control, logs, request limits, queue behavior, monitoring, backup of application state, and rollback when a new tag changes output.

For a shared endpoint, define the approved models and tasks. A smaller, faster model and a larger reasoning model can fill different deployment roles, but measured workload results—not a generic quality table—must justify the distinction.

Deployment model table

Model	Use	Required evidence
Developer workstation	Private experiments and task evaluation	Exact tag, local fit, data handling, repeatable prompt set
Single shared server	Small team endpoint	Access control, queueing, concurrency benchmark, monitoring, rollback
Containerized service	Repeatable runtime and application integration	Pinned image and model, persistent artifact storage, GPU mapping, health checks
Air-gapped environment	Workloads requiring network isolation	Artifact transfer, provenance, patching, logging, and update procedure
Hybrid routing	Local default with selected hosted fallbacks	Data classification, routing rules, measurable quality gate, audit trail

Migration checklist

Classify the workload, data sensitivity, and quality requirement.
Select the exact model tag and record its current artifact and model-card sources.
Review the checkpoint-specific license and base-model qualifications.
Define a representative prompt and evaluation set with expected outcomes.
Benchmark on target hardware with the model, quantization, context, and concurrency pinned.
Test Ollama request and response compatibility field by field.
Add access control, network restrictions, logs, health checks, and rollback.
Compare the measured local result with the current hosted alternative before cutover.

Source-backed evidence

Use Ollama artifact size and model-card context as screening inputs, not exact VRAM or throughput guarantees; benchmark on target hardware with pinned model, quantization, context and concurrency.

The Ollama DeepSeek library owns current tag and artifact information. The DeepSeek and Qwen model cards own checkpoint identity, license, parameter, context, and publisher-evaluation statements. Ollama's compatibility documentation owns the supported API surface and its limitations. None of those sources is a reproducible benchmark of your hardware or a current cloud-price comparison.

A reproducible benchmark worksheet

Input	Record before testing
Model	Exact tag, digest if available, and model-card revision
Runtime	Ollama version and host operating system
Quantization	Exact artifact or quantization label
Context	Configured context and actual prompt length
Hardware	CPU, GPU, memory, driver, and power mode
Concurrency	Number of parallel requests and queue policy
Tasks	Representative prompts with expected acceptance criteria
Outputs	Accuracy findings, latency distribution, failures, and resource use

This worksheet replaces unverified price, payback, throughput, concurrency, latency, and quality-ranking tables.

Methodology

OSSAlt reviewed five current primary sources on 2026-08-21: the Ollama DeepSeek-R1 and Qwen3 libraries, the DeepSeek-R1 and Qwen2.5-Coder-32B-Instruct model cards, and Ollama's API compatibility documentation. Publisher claims are labeled as such. Hardware fit, speed, latency, concurrency, cloud prices, and cross-model quality rankings were removed because no common dated benchmark or pricing packet was supplied.

Ollama DeepSeek-R1 library — accessed 2026-08-21
Ollama Qwen3 library — accessed 2026-08-21
DeepSeek-R1 official model card — accessed 2026-08-21
Qwen2.5-Coder-32B-Instruct official model card — accessed 2026-08-21
Ollama OpenAI compatibility — accessed 2026-08-21

Browse AI self-hosting guides at OSSAlt.