Modelpublic

Mantisec/Qwen3.5-4B-FP16

sourceHugging Faceupdated 21d agoView on Hugging Face

0likes152downloads

Model Card

Qwen3.5-4B-fp16

FP16 conversion of `Qwen/Qwen3.5-4B`, produced by bfsquish v0.1.0.

Intended use

Optimized for FP16 inference and fine-tuning on NVIDIA V100 (Volta) GPUs, which lack native BF16 Tensor Core support. The weights have been converted from BF16 to FP16 using the effective clamp_only strategy (see below).

Conversion details

Field	Value
Source model	`Qwen/Qwen3.5-4B`
Source revision	`latest`
Requested strategy	`auto`
Effective strategy	`clamp_only`
Tool	`bfsquish` v0.1.0
Converted at (UTC)	`2026-09-05T13:51:06.034581+00:00`
Target hardware	NVIDIA V100 (Volta, sm_70)
Target runtime	NVIDIA V100 (Volta sm_70), FP16 Tensor Cores

Conversion strategy

Direct upcast to FP32, clamp to FP16 range (+/-65504), downcast to FP16. The simplest strategy, near-lossless for well-behaved trained weights whose values are concentrated near zero.

Transform plan

v1 strategy applied: clamp_only (no v2 plan bundle was produced for this conversion).

Enhanced chat template

Prompt rendering uses the selected third-party enhanced template; model weights are unchanged by this step. The standalone and embedded forms were smoke-rendered to the same prompt before publication.

Field	Value
Selection	`latest`
Source	`peculiar-ragdoll/Qwen-Sharp-Chat-Templates`
Resolved revision	`fa3a1295882d31132770c156fced4e616b5db25d`
Template version	`qwen3.8-froggeric-v22.4.0`
Template license	`Apache-2.0`
Published forms	`chat_template.jinja` and `tokenizer_config.json#chat_template`

Reproduce the prompt-format step with --chat-template latest --chat-template-repo peculiar-ragdoll/Qwen-Sharp-Chat-Templates --chat-template-revision fa3a1295882d31132770c156fced4e616b5db25d.

Numerical quality

Metric	Value
Validation verdict	PASS
Validation method	`generate`
Max abs logit diff (vs. source)	0.215126
Min cosine similarity (vs. source)	0.999978
Token agreement rate	100.00%
Inf/NaN scan	passed (no inf/nan)

Validation notes

—generation agreement passed despite logit drift: tokenagreement=100.00% (pass≥98.00%), mincos=0.999978, max_diff=0.2151

Reproducing this conversion

bash

bfsquish run \
  --model Qwen/Qwen3.5-4B \
  --output-dir ./out \
  --strategy auto

License

Inherited from the source model. Refer to the source model's license for terms of use.