Can DeepSeek's New V4-Flash-0731 Run on Jetson Thor? The 284B MoE Edge Boundary
After checking Qwen3.8-Max, I looked at DeepSeek's recent open-weight releases. The model to watch is DeepSeek-V4-Flash-0731, not the older DeepSeek-V3 or R1 releases.
Short answer:
| Question | Answer |
|---|---|
| Recent DeepSeek open-weight model to watch | DeepSeek-V4-Flash-0731 |
| License | MIT |
| Scale | V4-Flash: 284B total / 13B activated |
| Context | 1M tokens |
| Can one Jetson Thor run it locally? | Not realistically |
| Practical test path | Run V4-Flash on a multi-GPU server; use Thor for 30B/35B local edge comparisons |
DeepSeek-V4-Flash-0731 is open and strong, but it is not a single-Jetson-Thor local model.
What Was Released
DeepSeek's Hugging Face organization shows deepseek-ai/DeepSeek-V4-Flash-0731 as a recent V4 release. The model card describes the V4-Flash family as:
| Item | DeepSeek-V4-Flash-0731 |
|---|---|
| Architecture | MoE |
| Parameters | 284B total / 13B activated |
| Hugging Face model page size | 304B params |
| Context length | 1M tokens |
| Precision | FP4 + FP8 mixed |
| License | MIT |
| Focus | code agents, tool use, long context, reasoning effort control |
DeepSeek-V4 also includes the larger DeepSeek-V4-Pro, listed as 1.6T total / 49B activated with the same 1M context target. Pro is the ceiling; Flash is the more efficient variant.
The 0731 checkpoint has major agent benchmark improvements over the preview release:
| Benchmark | V4-Flash-0731 | V4-Flash Preview |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 |
| NL2Repo | 54.2 | 39.4 |
| Cybergym | 76.7 | 38.7 |
| DeepSWE | 54.4 | 7.3 |
| Toolathlon-Verified | 70.3 | 49.7 |
| Agents' Last Exam | 25.2 | 15.8 |
This is clearly aimed at coding agents, tool use, and long-context work rather than simple chat.
Thor Check
I checked two live Thor machines:
| Device | Environment |
|---|---|
thor | NVIDIA Thor, CUDA 13.0, Linux 6.8.12-tegra, 122 GiB memory |
thor-jd | NVIDIA Thor, CUDA 13.0, Linux 6.8.12-tegra, 122 GiB memory |
Neither machine has DeepSeek or DSpark weights under /home/nvidia/models:
find /home/nvidia/models -maxdepth 2 -iname "*deepseek*" -o -iname "*dspark*"
So there is no local V4-Flash benchmark to report. More importantly, the model scale makes a single-Thor run the wrong target.
Why V4-Flash Does Not Fit One Thor
The phrase "13B activated" is easy to misread. It does not mean V4-Flash has the memory footprint of a 13B dense model.
V4-Flash is still a 284B-class MoE model. Even with FP4 experts and FP8 non-expert weights, it is far beyond Thor's comfortable 123 GB unified-memory envelope.
Rough weight-only estimates:
| Assumption | Size |
|---|---|
| 284B at FP8 | ~284 GB |
| 284B at INT4 | ~142 GB |
| FP4 + FP8 mixed | depends on non-expert ratio, but still near or above Thor's limit |
| 1M context KV cache | adds substantial memory pressure |
That excludes CUDA workspace, MoE routing buffers, prefill tensors, speculative decoding state, and server overhead.
The official model card's serving examples point to this hardware class. The vLLM example uses a single 4×GB300 node with data parallelism, expert parallelism, FP8 KV cache, and DSpark speculative decoding. The SGLang example uses --tp 4 and a specialized MoE backend:
sglang serve \
--model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
--tp 4 \
--moe-runner-backend flashinfer_mxfp4 \
--speculative-algorithm DSPARK
That is not a Jetson Thor deployment shape.
What Thor Should Test Instead
Use two layers:
| Layer | Test |
|---|---|
| Multi-GPU server | DeepSeek-V4-Flash-0731 agent quality |
| Jetson Thor | 30B/35B local edge models as a deployment baseline |
For V4-Flash, measure real agent tasks:
| Metric | Meaning |
|---|---|
| Pass@1 | Does it solve the task on the first attempt? |
| Total wall time | How long does the agent loop take? |
| Output tokens | How expensive is reasoning effort? |
| Tool-call count | Is it efficient or wandering? |
| Failure mode | command error, hallucination, format failure, missing verification |
For Thor, test models that actually fit:
| Model | Use case |
|---|---|
| Qwen3-Coder-30B-A3B-FP8 | local coding agent |
| Qwen3.6-35B-A3B-FP8 | local general agent / long-context assistant |
| Qwable-v1 | Claude-style agent behavior distillation |
| DeepSeek-R1-Distill-Qwen-14B / 32B | reasoning model if downloaded later |
This comparison answers the useful engineering question: how far is a local 30B/35B Thor agent from a frontier open-weight DeepSeek V4 agent?
Verdict
DeepSeek-V4-Flash-0731 is one of the most interesting recent open-weight models: MIT License, 284B MoE, 13B activated, 1M context, and much stronger agent benchmarks than the preview.
But it is still too large for a single Jetson Thor. The official examples already point toward TP4, expert parallelism, and high-end multi-GPU servers.
My practical choice:
- Test DeepSeek-V4-Flash-0731 on a multi-GPU server or hosted endpoint.
- Use Thor for Qwen3-Coder-30B-A3B-FP8, Qwen3.6-35B-A3B-FP8, Qwable-v1, or DeepSeek-R1 distilled models.
- Do not publish fake Thor token/s numbers for a model that is not locally present and not realistically supported by the hardware.
DeepSeek-V4-Flash-0731 is worth tracking, but Jetson Thor's role is a 30B/35B local edge baseline, not a 284B MoE single-node server.
Sources
- DeepSeek Hugging Face organization:
https://huggingface.co/deepseek-ai - DeepSeek-V4-Flash-0731 model card:
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 - DeepSeek-V4-Pro model card:
https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro - DeepSeek-V3 GitHub repository:
https://github.com/deepseek-ai/DeepSeek-V3
Thor environment checked on August 13, 2026. This post does not claim a local DeepSeek-V4-Flash-0731 benchmark because the model weights are not present on the Thor machines and the official serving path targets multi-GPU systems.