Unsloth’s 748bn-parameter coding model needs workstation-class memory

Unsloth’s new quantisation of DeepSeek’s 748-billion-parameter V4.1-Flash-1017 model reportedly scores 77% on the DeepSWE v1.1 coding benchmark, while its smallest practical build still needs substantial memory. This is a noteworthy step for open-weight coding models, though “runs locally” here does not mean “runs on your gaming PC”.

Unsloth Watch analysis

What happened

Startup Fortune reports that Unsloth used its Dynamic V3.1 method to quantise DeepSeek’s model, released in October. The publication says the 2-bit version needs about 176–205GB of memory, while a 1-bit version needs 128–145GB. That puts local use within reach of some workstation setups, not an ordinary single graphics card.

The article says the quantised model scored 77% on DeepSWE v1.1, compared with 74% for Claude Opus 5 and 73% for GPT-5.6-Sol. It also cites DeepSeek’s changelog, which records a rise from 31.2% to 42.3% on Terminal-Bench 4.0 between the earlier model and the updated checkpoint. Startup Fortune describes Dynamic V3.1 as a preview and says independent reproductions of the benchmark results are not yet available.

Why it matters

Quantisation reduces the memory needed to run large models, but the useful question is how much capability survives the squeeze. These reported results suggest Unsloth’s method may let developers run a large open-weight coding model on their own hardware with less loss than competing quantisation approaches.

The hardware bill remains a formidable footnote. A 128GB minimum is a very different proposition from downloading a model onto a typical developer’s computer, and benchmark scores alone do not establish how well it will perform on everyday coding work.

Our read

This is a serious local-inference development, not a new excuse to call every desktop a data centre. The reported scores are promising, and the memory requirements make the practical audience clear: teams with workstation-class systems, or those willing to rent one. Treat the benchmark lead as an interesting result to test, not a settled verdict.

What to watch

• Whether independent testing reproduces the DeepSWE and Terminal-Bench results.
• How Dynamic V3.1 performs on real coding tasks and different hardware configurations.
• Whether Unsloth moves the quantisation method beyond preview status.
• Whether the memory and hardware costs make local use competitive with API pricing.

Discussion spark: Would you take on workstation-class hardware to run an open coding model locally, or do API convenience and cost still win?

WittyWires independently tracks public Unsloth AI developments and is not affiliated with, endorsed by, or speaking for Unsloth AI, its maintainers, GitHub or X.