The upstream model is a 13B language model trained on 260B tokens of FineWeb. The original model card describes it as architecturally identical to talkie-lm/talkie-1930-13b-base and intended for controlled comparisons between vintage and modern language models.

The original base checkpoint is FP32. This repository stores a BF16 conversion of those weights and packages them for Transformers with custom trust_remote_code modules and BF16 sharded safetensors.

This is not an official Talkie release; refer to the upstream model card for the author-provided provenance and usage notes.

Source Model

Original model: talkie-lm/talkie-web-13b-base
Talkie report: talkie-lm.com
Reference code: github.com/talkie-lm/talkie

Conversion Details

Weight dtype: BF16
Weight format: sharded safetensors
Context length: 4096 tokens
Architecture: custom Talkie code loaded with trust_remote_code=True
Tokenizer: Talkie tiktoken-compatible tokenizer exposed through AutoTokenizer

The public reference configuration originally advertised 2,048 positions, but the Talkie team later clarified that the model was trained with a 4,096-token context. This conversion uses the corrected 4,096-token limit.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

path = "xlr8harder/talkie-web-13b-base-tf"
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    path,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map={"": "cuda"},
    use_safetensors=True,
)

For base completions:

python

inputs = tokenizer("The latest discoveries in physics suggest that", return_tensors="pt").to("cuda")
output = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(output[0], skip_special_tokens=True))

vLLM

The included remote-code model implements the Transformers attention-interface hooks expected by vLLM's Transformers modeling backend. For compatibility with that backend, the original single-scalar lm_head_gain is folded into lm_head.weight during conversion; the other Talkie gain parameters remain explicit model parameters. Using vLLM's logit_scale-style approach was not used because it applies scaling after the output matmul, while Talkie applies the gain to the head weight before the matmul. In BF16 this can introduce small rounding differences and, in smoke tests, changed one near-tied top-token ordering.

bash

vllm serve xlr8harder/talkie-web-13b-base-tf \
  --task generate \
  --model-impl transformers \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 4096

Validation

The Transformers safetensors model was compared against the original Talkie web FP32 checkpoint on a forward-pass smoke test. The top-10 next-token ordering matched exactly; observed max absolute logit difference was 0.03125.