Intel / repo

unsloth v0.1.909-beta — DeepSeek-V4.1 & FLUX-3

Source: github.com
Author: danielhanchen
Date: Published Oct 19, 2026

Context

The headline of this Unsloth release is local support for DeepSeek-V4.1-Flash-1017, which the release notes describe as an update to DeepSeek's V4.1-Flash with better reasoning, coding and agentic results and shorter outputs (GitHub). The useful detail is how the model fits on a single machine. According to the notes, it has 552B parameters plus another 196B in what the notes call its Engram table, with only about 16B of them active, and a 1M-token . With the Engram table left on SSD instead of in memory, Unsloth says the model runs in about 125 GB at 1-bit or about 300 GB at 4-bit, and it publishes GGUF files and a guide for doing so.

For background on what that table is, earlier SemiAnalysis coverage describes Engram as learned lookups for recurring multi-token patterns that extend ordinary token , an approach it argues lowers the high-bandwidth memory a model needs. That is the part of the model Unsloth's setup keeps on disk. The same release also adds Black Forest Labs' FLUX.3 Image Klein image model, multi-GPU fine-tuning with FSDP2 in alpha, and a one-command MCP setup for coding agents, but the memory figures above are what determine whether this DeepSeek model is runnable on a given workstation.

Key takeaways · AI-distilled

DeepSeek-V4.1-Flash-1017 has 552B parameters plus a 196B Engram table, about 16B active. Unsloth says keeping the Engram table on SSD lets it run in about 125 GB of memory at 1-bit or 300 GB at 4-bit.
Unsloth reports DeepSeek-V4.1-Flash-1017 at 42.3 on Terminal-Bench 4.0, up from 31.2 for the earlier V4.1-Flash. Thinking is on by default at high effort and can be set from low to max or turned off.
Fine-tuning gets two memory levers: optim='adamw_flash' cuts full fine-tuning optimizer memory, and UNSLOTH_OFFLOAD_OPTIMIZER_STATE=1 keeps optimizer state in system memory between steps at some cost in speed.
training with lora_dropout or lora_bias now runs on Unsloth's fused kernels, about 28% faster per step, and Gemma 4's global attention layers get a kernel with 1.4x to 2.1x faster attention and up to 4.9x less attention memory on Ampere or newer GPUs.
Through Unsloth's MCP server, coding agents can upload datasets, start training, load the resulting checkpoint, download models ahead of time and preview what a model load will download or unload.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.

Read the source github.com