Hero model · small reasoner

VibeThinker on Apple Silicon

A 1.5B and 3B reasoning model that punches well above its weight class — long chain-of-thought emitted as a structured reasoning_content field, suitable for distillation pipelines or low-RAM agents. rapid-mlx ships two MLX quants and the custom vibethinker reasoning parser.

TL;DR — one command. pip install rapid-mlx && rapid-mlx serve vibethinker-1.5b-4bit runs the 1.5B reasoner at 4-bit on a 4 GB-unified-memory Mac. The /v1/responses endpoint exposes the reasoning trace as a structured field — no string-parsing required. Headline tok/s: TODO — bench pending on M3 Ultra.

Why VibeThinker matters

The "reasoning at the low end" niche has been the gap between DeepSeek R1 (huge, server-only) and Qwen2.5-7B (small, but not reasoning-tuned). VibeThinker fills it — a 1.5B / 3B distilled reasoner that can show its work on math, code, and logic, while fitting comfortably in the RAM of an entry-level Apple Silicon Mac. If you need a chain-of-thought trace and you don't have 4 GB to spare for it, this is the model.

Engineering notes — what we shipped

The VibeThinker integration landed as a four-PR sequence, each addressing a different streaming-edge case in the reasoning parser:

The reasoning parser source is at vllm_mlx/reasoning/; the alias config at vllm_mlx/aliases.json.

Upstream status

Upstream mlx-lm can load VibeThinker weights but does not have a reasoning-parser surface — the <think>...</think> trace appears inline in the chat content. rapid-mlx splits the trace into a structured reasoning_content field on /v1/responses and on the OpenAI-style streaming delta, so clients don't have to string-match.

Pick your quant

Alias Bits Approx. disk Recommended for HF repo
vibethinker-1.5b-4bit 4 ~1.0 GB Default — fits 4 GB unified memory, including M1/M2 baseline VibeThinker-1.5B-mlx-4bit
vibethinker-3b-8bit 8 ~3.0 GB Quality bump — 8-bit 3B for distillation / long-CoT runs VibeThinker-3B-8bit

Position vs the big reasoners

DeepSeek R1, V4-Flash, GLM-Z1 — these are stronger reasoners, but they need a high-end workstation or a cloud GPU. Qwen3-Thinking and GLM-4.5-Air are middle-ground, requiring 24-32 GB of unified memory. VibeThinker is the only reasoning option in the ≤5 GB-RAM tier on Apple Silicon. If you need reasoning at all and your hardware budget is a 16 GB MacBook Air, this is the model. If you have 64 GB+, use DeepSeek V4-Flash instead — it is meaningfully stronger.

Recommended settings

The alias bakes in temperature=1.0, top_p=0.95 from the model card — VibeThinker was trained with high-temperature sampling for diversity of reasoning paths, and lower temperatures collapse the trace into shallow restatements. Tool calling uses the hermes parser. 

rapid-mlx serve vibethinker-1.5b-4bit \
  --port 8000 \
  --host 127.0.0.1

Tutorial — reasoning chain via /v1/responses

The structured reasoning trace shows up as response.reasoning — the message content is just the final answer, so you can render the trace separately or strip it for downstream evals. 

curl http://127.0.0.1:8000/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "vibethinker-1.5b-4bit",
    "input": "If a train leaves Chicago at 60 mph and another leaves NYC at 80 mph, when do they meet (790 miles apart)?",
    "reasoning": {"effort": "medium"}
  }'

Python with the OpenAI client:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="none")

resp = client.responses.create(
    model="vibethinker-1.5b-4bit",
    input="Prove that the sum of the first n odd numbers is n^2.",
    reasoning={"effort": "high"},
)

# Final answer
print(resp.output_text)

# Structured reasoning trace
for item in resp.output:
    if item.type == "reasoning":
        for chunk in item.summary:
            print("[reasoning]", chunk.text)

Performance

TODO — M-series headline tok/s pending. We'll publish 1.5B-4bit decode tok/s on M1 (8 GB), M2 Pro, and M3 Ultra in the next bench release.

Known limitations

Related