GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
GestaltLabs
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
GestaltLabs
Model Tree
deepseek-ai/DeepSeek-V4-Flash-0731
Input Modalities
Output Modalities
Supported Functionality
This checkpoint injects Ornstein thinking into DeepSeek-V4-Flash-0731. It is an early merge, not a finished quality release. Planned quality work uses RL environments and energy-based fine-tuning.
No downstream task evaluation or benchmark suite has been run on this checkpoint yet. Weight-level verification is not evidence of benchmark performance or parity with another model.
7872f01b1d1fe23eabc4c98b48bffcef5a386062), and 23 representative numerical probes confirmed BF16-rounding and FP32 tolerance against expected adapter updates.105ef75053e89b3700f8b6dc793fa7a8c04e12ee) containing model.safetensors.index.json and config.json. Independent download at that commit confirmed byte-for-byte metadata integrity and index completeness (36,599 tensors).I'm a PhD student in visual neuroscience at the University of Toronto. Training and release compute is self-funded (rented H100s and a local DGX Spark). If these artifacts are useful, Ko-fi helps keep the experiments running.
| Attribute | Specification |
|---|---|
| Architecture | DeepseekV4ForCausalLM (deepseek_v4) |
| Layers | 43 hidden layers (3 hash-routing layers, num_hash_layers: 3) |
| Hidden size | 4096 |
| Attention | 64 query heads, 1 KV head (Multi-Head Latent Attention), head_dim 512 |
| Sliding window | 128 tokens |
| Mixture of Experts | 256 routed experts (moe_intermediate_size 2048), 6 active experts per token, 1 shared expert |
BF16 safetensors storage does not itself establish runtime compatibility:
transformers==4.57.6) lacks support for the DeepseekV4ForCausalLM architecture. Upstream development source (transformers commit cbc1651a, identifying as version 5.17.0.dev0) implements a source-level main trunk mapping with eager attention, but explicitly ignores mtp.* speculative decoding weights and disables assisted generation; full BF16 + DSpark runtime execution has not been verified.The upstream tokenizer does not provide a Jinja chat template. Use the included encoding/encoding_dsv4.py for DeepSeek-V4 prompt formatting and response parsing. Run this example from the downloaded repository directory:
python
from encoding.encoding_dsv4 import encode_messagesmessages = [{"role": "user", "content": "Explain the concept of Gestalt grouping principles in visual neuroscience."}]# Formats conversation into the required model promptprompt = encode_messages(messages,thinking_mode="thinking",reasoning_effort="high" # Supported: "low", "high", "max")
This encoding example and a tokenizer encode/decode round trip were exercised successfully without loading model weights. That is not a full inference test.
Download the full repository using the Hugging Face CLI:
bash
# Requires about 610 GB for the files, plus download overhead.# Full BF16 inference requires a suitable multi-accelerator system.hf download GestaltLabs/DeepSeek-V4-Flash-0731-Ornstein --local-dir ./DeepSeek-V4-Flash-0731-Ornstein
| Path | Description |
|---|---|
model-00001-of-00048.safetensors … 00048 | Dequantized BF16 model shards (48 shards, ~608.5 GB target payload) |
model.safetensors.index.json | Shard tensor weight mapping index |
config.json | Architecture configuration (DeepseekV4ForCausalLM) |
generation_config.json | Generation hyperparameters |
tokenizer.json / |
MIT License, inherited from the upstream DeepSeek-V4-Flash-0731 release.
| Context window | 1,048,576 tokens (YaRN RoPE scaling: factor 16, original_max_position_embeddings 65536) |
| Vocabulary size | 129,280 tokens |
| Multi-Token Prediction (MTP) / DSpark | 3 physical DSpark stages (mtp.0, mtp.1, mtp.2; target layers [40, 41, 42], dspark_block_size: 5, dspark_markov_rank: 256; base config records num_nextn_predict_layers: 1). All upstream MTP tensors are preserved and dequantized to BF16 (no LoRA applied to MTP) |
| Target Release Payload | 36,599 tensors across 48 shards: 36,163 BF16, 433 preserved F32, 3 preserved I64 controls (608,450,282,488 payload bytes; 2,350 paired UE8M0 scales removed; 24 shards direct, 24 rewritten) |
| Post-training | PEFT LoRA trained on Fireworks AI; merged across 33,450 linear weight tensors |
tokenizer_config.json| DeepSeek-V4 tokenizer files |
encoding/ | Upstream OpenAI message encoding scripts (encoding_dsv4.py) |
LICENSE | MIT License |