Private-codebase specialization
Train an agent on internal APIs, conventions, and repositories while retaining control of the data and resulting weights.
Independent tool overview
SERA at a glance
SERA is Ai2's open family of repository-level coding agents, released with model weights, training data, specialization code, and a Claude Code-compatible CLI.
Visit the official SERA site ↗
- Developer
- Ai2
- Full name
- Soft-Verified Efficient Repository Agents
- Current sizes
- 8B, 14B, and 32B
- License
- Apache 2.0 with Ai2 Responsible Use Guidelines
- Primary interface
- Claude Code-compatible SERA CLI or vLLM endpoint
- Context used in published evaluations
- 32K
Overview
What SERA is
SERA stands for Soft-Verified Efficient Repository Agents. It is a family of open coding models from Ai2 designed for repository-level tasks such as debugging, code generation, review, maintenance, and explanation. Current releases include 8B, 14B, and 32B variants, plus variants trained from an earlier teacher model.
The more distinctive part of SERA is the reproducible data-generation and fine-tuning recipe. Teams can generate agent trajectories from their own repositories and specialize a smaller open model to internal APIs, conventions, and code patterns while controlling where the code and model run. Ai2 also provides a proxy that can launch SERA behind Claude Code through Modal, an existing vLLM endpoint, or self-hosted inference.
Use cases
Who SERA is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Coding-agent research
Inspect the released models, datasets, generation pipeline, verification metadata, and evaluation recipe.
Self-hosted coding assistance
Run an open repository agent behind a private vLLM endpoint instead of sending code to a closed hosted model.
Claude Code interface experiments
Use the familiar Claude Code client with a SERA model through Ai2's compatibility proxy.
Cost-controlled fine-tuning studies
Explore synthetic data and supervised specialization without building a large reinforcement-learning stack from scratch.
Capabilities
Core SERA features
Open model family
Provides downloadable SERA-8B, SERA-14B, and SERA-32B weights, including additional teacher variants.
Repository-level agent behavior
Models are trained for multi-step software tasks that require inspecting, editing, and testing code across a repository.
Private-repository specialization
The released pipeline can generate training trajectories from arbitrary personal or internal codebases.
Training recipe and data
Ai2 releases generation code, configuration, datasets, metadata, and training instructions rather than only model weights.
Agent-framework support
The data pipeline documents support for SWE-agent and mini-swe-agent workflows.
Claude Code compatibility
The SERA CLI runs a proxy so Claude Code can use a SERA model served through vLLM.
Ephemeral Modal launch
A single CLI route can provision the model on Modal for one session and clean up the app when Claude Code exits.
Persistent team deployment
A separate command can create a shared long-running SERA endpoint with an automatically generated API key.
Self-hosted inference
Teams can serve SERA directly with vLLM on their own GPU infrastructure and point the proxy at that endpoint.
Model-agnostic refreshed datasets
Ai2's release update adds verification thresholds and metadata intended to make filtering and downstream training easier.
Process
How the SERA workflow works
Step 1
Choose the objective
Decide whether the need is evaluation, daily coding assistance, research reproduction, or specialization to a private repository.
Step 2
Select a model size
Balance benchmark results, latency, GPU memory, context needs, and concurrency across the 8B, 14B, and 32B variants.
Step 3
Isolate the runtime
Use a sandboxed development environment with scoped repository access, short-lived credentials, resource limits, and no production secrets.
Step 4
Launch an endpoint
Use the SERA CLI with Modal, connect it to an existing endpoint, or host the model through vLLM.
Step 5
Evaluate on local tasks
Test representative issues from the target codebase and measure correct patches, test results, regressions, security findings, latency, and cost.
Step 6
Specialize if justified
Generate training data from approved repositories, filter and verify samples, fine-tune a selected model, and retain a clean holdout evaluation set.
Step 7
Keep humans in the merge path
Require code review, tests, security scans, and explicit approval before any SERA-generated change reaches a protected branch or deployment.
Cost
SERA pricing and free plan
SERA's model weights, CLI, and training code are available without a software subscription under open licenses, but running and fine-tuning the models requires GPU infrastructure. Ai2's published dollar figures are examples from its experiments, not fixed plans or guaranteed budgets.
Models and code
Free to download
Open SERA weights, training repository, datasets, and CLI.
- Model cards list Apache 2.0 licensing
- Use remains subject to applicable license terms and Ai2 Responsible Use Guidelines
- Storage, bandwidth, compute, and operations are not included
Self-hosted inference
Infrastructure cost
Run SERA on owned or rented GPUs through vLLM.
- Cost depends on model size, GPU, context length, utilization, and concurrency
- The 14B model card lists one 80GB GPU as a minimum for its documented 32K setup
- Quantization can reduce memory needs with possible quality tradeoffs
Modal deployment
Usage-based cloud cost
The CLI can provision ephemeral or persistent SERA endpoints on Modal.
- Requires a Modal account
- Charges depend on provisioned GPU and runtime
- Ephemeral mode cleans up when the local session exits
- Persistent deployments must be stopped when no longer needed
Repository specialization
Variable training and teacher cost
Generate synthetic trajectories and fine-tune on a target codebase.
- Ai2 reports about $400 in one reproduction scenario
- Ai2 reports up to $12,000 in a higher-performance scenario
- A separate 8,000-sample specialization example is reported at $1,300
- These figures are research examples, not quotes or forecasts
Pricing checked . Check current pricing at the source ↗
Assessment
SERA strengths and limitations
Where it stands out
- Releases the full stack of weights, data, generation code, training recipe, and integration tooling
- Makes codebase-specific specialization accessible to teams that cannot build a large RL pipeline
- Can keep private source and inference inside infrastructure controlled by the user
- Offers several model sizes for different performance and hardware constraints
- Claude Code compatibility provides a practical interface for testing the models
- Model and pipeline licenses reduce software lock-in
- Published datasets and configurations support reproducible research and inspection
What to consider
- SERA is a research-oriented model family, not a fully managed coding service with hosting, support, billing, and enterprise controls
- Official model cards say the models have no safety filtering or safety tuning and require significant human oversight
- Models can generate incorrect, insecure, destructive, or subtly regressive code
- Repository access creates prompt-injection and data-leakage risks through source files, issues, tests, tool output, and dependencies
- The strongest models require substantial GPU memory and operational expertise
- SWE-bench results do not predict reliability on a specific internal repository, language, build system, or deployment environment
- Fine-tuning on private code requires careful data rights, secret removal, licensing review, evaluation design, and model-access governance
- Claude Code compatibility does not mean SERA has the same behavior, safety, or quality as Anthropic's Claude models
- Cloud convenience through Modal still means code context and model traffic leave local hardware unless the environment is configured and approved appropriately
Compare
SERA alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Devstral 2
Mistral's open agentic coding model family offers another self-hostable route with its own CLI ecosystem.
Explore Devstral 2 →Coding
Claude Code
A managed coding-agent experience for teams prioritizing frontier model quality and a supported end-user workflow over open weights.
Explore Claude Code →Coding
Gemini CLI
Google's open-source terminal agent offers a lower-friction hosted-model workflow with high free usage allowances.
Explore Gemini CLI →Questions
SERA FAQs
What is SERA?
SERA is Ai2's family of Soft-Verified Efficient Repository Agents: open coding models and a training recipe for repository-level software tasks.
Which SERA models are available?
Current model cards list SERA variants based on 8B, 14B, and 32B Qwen 3 models, plus variants trained from an earlier GLM teacher.
Is SERA free?
The weights, code, and CLI are free to download under their listed licenses. Users still pay for storage, GPU inference, cloud hosting, teacher-model calls, and any fine-tuning work.
Can SERA run in Claude Code?
Yes. Ai2's SERA CLI provides a compatibility proxy that lets Claude Code use a SERA model served on Modal, an existing endpoint, or self-hosted vLLM.
Can SERA learn a private codebase?
The released pipeline supports generating training trajectories from arbitrary repositories and fine-tuning a model to those patterns. Organizations must still protect secrets, verify data rights, and evaluate the specialized model independently.
What hardware does SERA need?
Requirements depend on model size, context, precision, and concurrency. The SERA-14B model card describes one 80GB A100 or H100 for its documented 32K setup, while the 32B weights and high-throughput deployments require more capacity.
How good is SERA on SWE-bench Verified?
The current model card reports 49.5% plus or minus 1.9 for SERA-32B, 41.7% plus or minus 0.5 for SERA-14B, and 31.7% plus or minus 0.9 for SERA-8B at 32K context across three seeds. These are benchmark results, not guarantees on real repositories.
Is SERA safe for autonomous production changes?
No. Ai2's own model cards describe the models as unsafety-tuned research artifacts. Use sandboxing, least privilege, tests, security review, and mandatory human approval.
Bottom line
Our SERA verdict
SERA is one of the more complete open releases for teams that want to study or specialize a repository agent instead of treating coding AI as a closed API. Its openness is the opportunity and the burden: users gain control over data and weights, but also inherit GPU operations, evaluation, sandboxing, and the safety work that a managed service would normally absorb.
Visit SERA website ↗Similar Coding Tools
View all Coding AI tools →
NousCoder-14B
NousCoder-14B - Nous Research's competitive olympiad programming model
Codex App
Codex App - OpenAI's new Mac app interface for managing agents with its Codex agentic coding assistant
GLM-4.7
GLM-4.7 - Z AI's new SOTA open-source model with advanced coding and reasoning capabilities
Qwen3-Coder-Next
Qwen3-Coder-Next - Alibaba's open-weight, small hybrid model for agentic coding
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.