Lissa1-4B

A small senior-developer model for developers who want to ship.

Lissa1-4B is the text-only Lissa1 developer release built on the Qwen3.5-4B family and tuned for practical software work: reading a repository, explaining a trade-off, writing a patch, debugging a failure, and following a structured tool workflow.

This is not marketed as a giant general-purpose assistant. It is deliberately a compact developer model: small enough to run locally, useful enough to sit inside an IDE, terminal agent, CI helper, or private coding service.

Developer-to-developer notes

The intended user is another developer. Treat the model as a fast junior-to-mini-senior pair programmer, not as an authority:

• give it the repository context and the exact error;
• ask for a plan before a destructive change;
• make it show commands and verification steps;
• run the generated code and tests yourself;
• keep temperature low for patches and tool calls;
• do not let a fluent answer replace an execution result.

The model is text-only. It does not need an image processor, vision tower, or multimodal prompt.

Quick start

pip install -U transformers torch accelerate

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "lomelidev/Lissa1-4B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{
    "role": "user",
    "content": "Review this function for a race condition and propose a minimal fix."
}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        inputs,
        max_new_tokens=512,
        do_sample=False,
    )
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

Sampling presets

These are safe starting points, not a claim that one setting is optimal for every codebase.

Workload	temperature	top_p	max tokens	Notes
Patch/tool call	0.0–0.2	0.90–1.0	512–2048	Prefer deterministic output and validate every call.
Debug explanation	0.2–0.4	0.90	512–1536	Good default for developer chat.
Architecture alternatives	0.4–0.6	0.90–0.95	1024–2048	Ask for explicit trade-offs.
Brainstorming	0.6–0.8	0.95	1024–2048	Re-run the useful ideas deterministically.

For evaluation, use temperature=0, do_sample=False, a fixed prompt file, and record the tokenizer revision and output tokens.

What to verify

A minimum smoke test should check:

1. the tokenizer loads;
2. the model produces non-empty text;
3. the chat template adds a generation prompt;
4. the output stops at the expected EOS;
5. a small Python or shell example actually runs;
6. a requested tool call validates against the real schema.

The model is not a substitute for tests, static analysis, a database, or a deployment review. It can be confidently wrong, especially when repository context is incomplete.

Format repositories

Use the same model through the format that matches your runtime:

• GGUF / llama.cpp — Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q4_K_S and IQ4_XS.
• FP8 / vLLM and SGLang — the measured GPU serving release.
• AWQ / vLLM and SGLang — practical 4-bit serving.
• MLX / Apple Silicon — F16, 8-bit, 6-bit and 4-bit.
• DFlash drafter — speculative decoding sidecar, not a standalone model.

The complete family is collected in Lissa1.

Long-context test provenance

The repository's long-context test design is based on the supplied ia-lab Ornith-1.5-9B report: explicit slots, separate KV quantization, Flash Attention, continuous batching, YaRN and a real agent battery. Those measurements are for Ornith, not Lissa1. Lissa1 cards will only receive 4B numbers after reproducing them with the Lissa checkpoint and recording engine/version, VRAM, prompt tokens, prefill, decode and correctness.

One important non-transfer: Ornith's vLLM MTP path reached 90.09% acceptance, but Lissa1-4B has no MTP tensors after the merge. The 4B path is DFlash/community-DFlash, not MTP.

QLoRA roadmap

The base model is the root for planned, separate developer adapters. These are planned artifacts, not repositories that already exist:

• database: PostgreSQL, Redis, MongoDB, SQL review and query plans;
• frontend: TypeScript, React, Next.js, accessibility and UI testing;
• backend: APIs, Go services, Node.js and integration tests;
• devops-cicd: Docker, Kubernetes, K3s, Terraform, GitHub Actions and incident response;
• agents-tools: tool schemas, MCP, terminal loops and safe execution;
• security: secure coding, dependency review and threat modeling;
• platform: Proxmox, Talos, networking, observability and operations.

Each adapter will declare base_model: lomelidev/Lissa1-4B, publish its own evaluation, and be accepted only after a regression check against the global model. Adapters will not be silently merged into this base release.

Provenance and limitations

Lissa1 is a developer-oriented fine-tuned release. Upstream Qwen3.5 licensing remains applicable. Check the upstream license before redistribution. Do not use this model for autonomous production changes without a human gate, tests, and rollback.