Local candidates
Start with a model that fits the task, evaluation and hardware.
Open-weight models and local fine-tuning
Adapt open models. Keep the evidence.
Project PULSAR treats model selection and adaptation as reproducible engineering: choose by workload, tune only when evidence supports it, and keep data rights, evaluations and rollback attached to every release.
Frontier references
Use large open-weight systems as measured reference points.
Controlled tuning
Trace data, adapters, evaluations and releases end to end.
Reproducible research
Pin artifacts, publish assumptions and preserve rollback.
Sovereign boundaryAdaptive routingPermission-aware contextHuman authority
01 / Open-weight portfolio
A portfolio, not a leaderboard
The strongest model is the smallest governed option that passes the workload’s quality, latency, risk and cost gates.
LOCAL ADAPTOfficial model card
Qwen3-8B
Practical local fine-tuning candidate
- Scale
- 8.2B parameters
- Context
- 32K native · 131K with YaRN
- License
- Apache 2.0
A useful starting point for domain behavior, structured output and bilingual experiments. “Qwen 3.8” is interpreted here as the official Qwen3-8B model.
CLUSTER PILOTOfficial model card
Qwen3.6-35B-A3B
Efficient medium-scale MoE candidate
- Scale
- 35B total · 3B activated
- Context
- 262K native
- License
- Apache 2.0
A current open-weight MoE reference for evaluating quality, latency and memory trade-offs on a capable local server.
FRONTIER REFERENCEOfficial model card
Kimi-K2.6
General and multimodal reference
- Scale
- 1T total · 32B activated
- Context
- 256K
- License
- Modified MIT
A frontier-scale open-weight system for architecture and evaluation research. Its storage and serving envelope makes it a cluster reference, not a first fine-tuning target.
FRONTIER REFERENCEOfficial model card
Kimi-K2.7-Code
Agentic coding reference
- Scale
- 1T total · 32B activated
- Context
- 256K
- License
- Modified MIT
A coding-focused open-weight reference for long-horizon software tasks, tool use and controlled agent evaluation.
FRONTIER REFERENCEOfficial model card
Qwen3.5-397B-A17B
Large multimodal reference
- Scale
- 397B total · 17B activated
- Context
- 262K native
- License
- Apache 2.0
A large MoE reference for multimodal and agentic evaluation. It belongs in the portfolio as a benchmark target, not as the default local adaptation model.
FRONTIER REFERENCEOfficial model card
GLM-5
Long-horizon systems reference
- Scale
- 744B total · 40B activated
- Context
- 200K
- License
- MIT
A large open-weight reference for complex systems engineering and long-horizon agent research. Validate the exact workload instead of relying on leaderboard rank.
02 / Adaptation decision
Tune only the part that needs to change
Fine-tuning is one instrument in the system. Retrieval, tools and policy often solve the problem with less risk.
Prompt and tools
Start here when the model already has the capability and the missing piece is instruction, workflow or deterministic computation.
Retrieval-augmented generation
Use governed retrieval for changing facts, private knowledge and source-backed answers. Fine-tuning is not a knowledge database.
LoRA
Train small adapter matrices when behavior, tone, format or domain task performance needs repeatable adaptation.
QLoRA
Use a quantized base model plus trainable adapters when local accelerator memory is the primary constraint.
Continued pre-training or full tuning
Reserve for evidence-backed cases with sufficient data rights, compute, evaluation depth and recovery controls.
03 / Unsloth local laboratory
From private dataset to governed adapter
Unsloth provides local LoRA and QLoRA workflows for Qwen and other open-weight families. PULSAR wraps the training step in evidence, data rights and release control.
01
Charter the data
Record ownership, consent, license, purpose, retention and prohibited content before a training row exists.
02
Freeze the baseline
Pin model revision and checksums; create immutable quality, safety and memorization evaluations.
03
Run an isolated experiment
Use Unsloth with a pinned environment, LoRA or QLoRA, limited access and complete experiment metadata.
04
Challenge the result
Compare held-out quality, refusal, privacy, bias and task regressions against the untouched baseline.
05
Package provenance
Version the adapter, tokenizer, prompt template, dataset manifest, evaluation report and model card together.
06
Release with rollback
Canary the bounded use case, monitor real outcomes and preserve a tested route back to the previous artifact.
RESEARCH STARTER
A minimal QLoRA experiment
Use the current official Unsloth notebook, pin package and model revisions, and replace the example dataset with a rights-cleared, versioned dataset. The snippet is a starting shape—not a production recipe.
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="unsloth/Qwen3-8B",
max_seq_length=2048,
load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
model,
r=16,
target_modules=[
"q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",
],
)04 / Hardware envelopes
Match the experiment to the memory budget
Vendor-reported fit figures are planning references. Measure the exact sequence length, batch, optimizer, precision and checkpoint strategy before procurement.
Local QLoRA laboratory
Qwen3-8B and smaller Qwen3.5 variants are practical starting points. Unsloth reports Qwen3-14B fitting a 16 GB T4; actual fit still depends on sequence length, batch, optimizer and checkpointing.
Deeper adaptation pilot
Explore larger dense models or 30–35B MoE adapters. Unsloth reports about 56 GB for Qwen3.5-27B bf16 LoRA and 74 GB for Qwen3.5-35B-A3B bf16 LoRA.
Frontier serving and evaluation
Kimi, large Qwen and GLM references need substantially more accelerator memory, storage, networking and operations. Measure them as shared infrastructure, not workstation projects.
Memory figures attributed to Unsloth are vendor-reported examples, not PULSAR benchmark results. Recalculate infrastructure cost and fit on the date of each experiment. Method source ↗
05 / Release gate
A trained model is not yet a trusted model
Open weights do not automatically mean open source, unrestricted commercial use or acceptable risk. Every candidate must pass license, data, evaluation and operational review.
- Remove secrets and direct identifiers before training.
- Keep train, validation and final evaluation sets separate.
- Review the model license and dataset rights for the intended use.
- Test memorization, unsafe completion and permission-boundary regressions.
- Register every adapter and retain its full provenance.
- Require a named human owner and a tested rollback path.
Manifesto & rationale
The thinking behind the system
Why model choice is a portfolio decision
There is no universally strongest model. A model is strong for an organization when it satisfies the actual workload’s quality, latency, language, modality, license, privacy and operating-cost requirements inside an enforceable trust boundary.
Project PULSAR therefore separates local adaptation candidates from frontier references. A compact model may be the right system for a high-volume structured task; a frontier-scale mixture-of-experts model may remain a shared evaluation target until its infrastructure envelope is justified.
What fine-tuning should change
Fine-tuning can make behavior more consistent: task format, domain terminology, response structure, classification boundaries or tool-calling patterns. It should not replace permission-aware retrieval for changing facts or private records.
The safest progression is prompt and tool design, governed retrieval, LoRA, QLoRA and only then more invasive training. Each step must outperform the simpler baseline on held-out quality and safety evaluations.
The PULSAR research position
Open weights create valuable freedom to inspect, evaluate, host and adapt model artifacts. They do not remove licensing duties, dataset rights, privacy risk, operational ownership or the need for human authority.
Model names and availability are also versioned evidence. This page records its verification date and links to official model cards. A future Kimi K3 release will enter the reference set only after official weights, model card, license and checksums can be independently verified.