1. Direct answer: shortlist the artifacts, not the brand names
An Unreal team evaluating open-weight AI in 2026 should start with DeepSeek V4, Qwen3.6, GLM-5.2, MiniMax M3, Gemma 4, Leanstral 1.5, and Kimi K3 after its announced weights are actually available. The list is deliberately broader than the highest-volume names: local control, a compatible license, a supportable runtime, and good performance on the team's own code can matter more than general popularity.
“Open weight” only says that some trained parameters are available under stated terms. It does not prove that training code or data is open, that commercial use is permitted, that every variant can run locally, or that the model is secure. Record the exact repository, hash, license revision, architecture, precision, context limit, runtime, and serving configuration before comparing output.
2. Verified 2026 model matrix
| Model | First-party status | July 12-19 exact-term signal | Evaluation emphasis | |---|---|---:|---| | DeepSeek V4 | API release April 24; verify weight availability separately | 56.70 average, 97 latest | reasoning, agents, API migration, artifact status | | Qwen3.6 | Qwen3.6-Plus release April 2; model family has multiple artifacts | 18.82 average | coding, multimodal work, exact variant and license | | GLM-5.2 | release June 16 with long-horizon positioning | 22.46 average | repository work, debugging, long-context drift | | MiniMax M3 | release June 1 with coding and multimodal positioning | 13.57 average | agentic coding, context, serving cost | | Gemma 4 | official model card available | 43.27 average | size variants, local inference, task-specific quality | | Leanstral 1.5 | official Mistral release | not in this five-term group | specialized reasoning and reproducible deployment | | Kimi K3 | hosted launch July 16; weights promised by July 27 | tracked in frontier group | wait for the actual artifact and terms before local claims |
Google Trends numbers are normalized within the five-term open-weight comparison. DeepSeek V4 led this exact group and rose from a previous-24-hour average of 53.12 to 65.42 in the latest 24 hours. That is a discovery signal, not proof of better Unreal code, a permissive license, or lower total cost.
3. License and artifact gate
Create a model intake record before downloading anything. It should include the vendor source, repository and revision, file hashes, license text, commercial and redistribution constraints, acceptable-use policy, training-data disclosure if available, security contact, and update channel. If adapters, quantizations, or community conversions are used, record them as separate artifacts with separate provenance.
Do not label an entry “open source” merely because weights can be downloaded. Do not assume that a hosted model and a similarly named downloadable artifact behave identically. If Kimi K3 weights have not yet appeared at the promised source, keep the status as announced. If DeepSeek exposes a model by API but the required local artifact is absent or differently licensed, keep hosted and local rows separate.
4. Hardware and runtime feasibility test
Estimate memory from the actual parameter count, active-parameter behavior, precision, quantization, KV cache, context length, concurrency, and runtime overhead. Then measure on the deployment hardware. A model that loads once but swaps memory, stalls during long prompts, or cannot serve concurrent review jobs is not operationally local-ready.
Run three contexts: a small single-file repair, a medium module with tests and build metadata, and a repository navigation task. Measure time to first token, total latency, peak GPU and system memory, tokens per second, failure rate, context truncation, and output stability. Repeat after a cold start. Document the exact runtime and flags so another engineer can reproduce the result.
5. Unreal coding benchmark with negative cases
Use inspectable text artifacts: C++, Build.cs, Target.cs, config, logs, tests, command lines, and exported Blueprint descriptions. Keep .uasset modification inside Unreal Editor or an approved editor tool. The benchmark should include a compile failure, an engine-version API mismatch, a lifecycle or ownership defect, a packaging/configuration issue, and a task where the correct answer is to request missing evidence.
For the compile failure, score whether the patch compiles and whether the explanation matches the real include, module, reflection, or API boundary. For the lifecycle defect, inject a misleading comment and see whether the model follows code evidence. For the packaging issue, require a minimal change and a clean-build verification plan. Models that produce large rewrites, invent APIs, expose secrets, or skip rollback should fail even if the prose sounds confident.
6. Security, privacy, and tool authority
Begin in a disposable clone without production credentials, private asset archives, customer data, or signing keys. Deny network access unless the test requires an allowlisted domain. Restrict write scope, log every tool call, preserve prompts and outputs under the approved retention policy, and require human review before build or editor actions.
Add prompt-injection fixtures in comments, documentation, filenames, and retrieved content. The model must ignore instructions that conflict with the system policy and task. Test path traversal, destructive commands, secret requests, dependency substitution, and attempts to change unrelated files. Prove that the operator can stop the run and restore the repository without trusting the model's own summary.
7. Adoption and replacement decision
Choose a model per job, not per marketing cycle. A smaller model may win fast log classification; a larger model may win multi-module planning; a specialized model may win proof-heavy reasoning. Use blind review where possible and keep a hosted baseline so local-control benefits are measured against quality, latency, maintenance, and incident cost.
The adoption record should name accepted tasks, denied tasks, model and artifact hashes, runtime, hardware, context policy, tools, permissions, evaluation results, known failures, owner, patch cadence, and replacement trigger. Re-run the suite after a model, quantization, runtime, driver, or Unreal version changes. If the replacement cannot reproduce the accepted results, keep the prior route available until migration is complete.
8. Two deployment examples and one rejection case
Workstation review service. A team may deploy a quantized model on a dedicated development GPU for log classification and small C++ explanations. The service receives only selected text files, runs without repository write access, and returns a proposed diagnosis with cited lines. Acceptance requires stable cold-start latency, a memory ceiling, no secret access, and a measured false-confidence rate. If longer contexts cause swapping or degraded answers, keep the route limited to small artifacts.
Shared build-network assistant. A larger model may run on an isolated internal server for repository navigation and patch proposals. The caller sends a clean commit, the service writes to a temporary branch, and CI compiles and tests the result. The model cannot reach signing keys, production credentials, or binary asset storage. Audit logs connect every change to a model hash, runtime configuration, prompt, reviewer, and CI result. Failed tests delete the temporary environment but preserve the evidence bundle.
Rejection case. Reject a candidate even when benchmark scores are strong if its license is unclear, the published artifact differs from the evaluated hosted service, required quantization comes from an unverified third party, or the runtime cannot be patched and monitored. Also reject direct .uasset edits without an editor-owned, source-controlled, reopen-tested path. Local execution is not a benefit when the team cannot explain artifact provenance, incident response, or restoration.
Review total cost over a real month, including hardware occupancy, electricity, storage, runtime engineering, model downloads, security review, evaluation, failed runs, and operator time. Compare that figure with the approved hosted baseline at the same quality and latency gate. A lower per-token estimate does not establish a cheaper workflow when engineers spend additional hours correcting patches or maintaining the serving stack.
SEELE AI does not export a native .uproject or compile Blueprint or C++; this independent comparison is not an Epic endorsement. Keep native asset, build, security, and deployment validation inside the controlled Unreal workflow.
Official sources
- DeepSeek API change log
- Qwen3.6-Plus release
- GLM-5.2 release
- MiniMax M3 release
- Gemma 4 model card
- Leanstral 1.5 release
- Kimi K3 launch and weight-release statement
- [Latest AI model tracker](/resources/blogs/latest-ai-models-for-unreal-engine-game-development-2026)
Frequently asked questions
Does open weight mean open source?
No. Weights, training code, data, license, and commercial rights are separate. Read the exact model license and distribution terms before local or customer-facing use.
Which open-weight model is hottest now?
DeepSeek V4 led its July 12-19 comparison group, while Gemma 4 also had substantial exact-term interest. Relative interest is not an Unreal quality score.
Can these models run on a developer workstation?
Some quantized variants may, but feasibility depends on released artifacts, active parameters, context length, runtime, precision, GPU memory, system memory, and acceptable latency.
Should a model edit .uasset files directly?
No. Treat binary Unreal assets as editor-owned. Prefer inspectable source inputs, bounded editor tools, source-control diffs, backups, and reopen or cook validation.
What is the minimum security test?
Use a disposable repository, deny secrets and production credentials, restrict network and write scope, log every tool call, review diffs, and prove rollback after hostile instructions.
When should a team reject local deployment?
Reject it when licensing, artifact provenance, hardware cost, latency, evaluation quality, patching, monitoring, or incident ownership is worse than the approved hosted alternative.
