Unsloth Brings Mistral Medium 3.5 to Any PC With 64GB RAM
Unsloth releases quantized GGUFs of Mistral's new 128B unified model, making a frontier-class reasoning, coding, and vision model runnable on ~64GB RAM
PRO- Unsloth releases quantized GGUFs of Mistral Medium 3.5, a new 128B dense model, runnable locally on ~64GB RAM using Unsloth's Dynamic 2.0 method.
- One model replaces three: Mistral Medium 3.5 unifies Magistral (reasoning), Devstral 2 (coding), and Mistral Medium (chat) into a single set of weights.
- Key specs: 128B parameters, 256k context window, multimodal input (text + images), configurable reasoning effort per request.
- Benchmarks: 77.6% on SWE-Bench Verified and 91.4% on τ³-Telecom; API priced at $1.50/M input tokens, roughly half the cost of Claude Sonnet.
- Unsloth bug fix included: A YaRN parsing quirk causing models to forget conversations was identified and fixed in collaboration with Mistral.
- Available now via Hugging Face GGUFs, Mistral API, Le Chat, and NVIDIA NIM; vision in GGUF format is not yet supported.
Unsloth has published quantized GGUF versions of Mistral's newly released Mistral Medium 3.5, a 128B dense model that collapses what used to be three separate Mistral products into one. The GGUFs, built using Unsloth's Dynamic 2.0 quantization method, bring a frontier-class model within reach of anyone with ~64GB of RAM and no data center required.
One model to replace three
Mistral has been shipping specialized model lines for two years: Magistral was the reasoning specialist, Devstral 2 was the coding specialist, and the base Mistral Medium series was the chat workhorse, each with its own training run, tuning, and integration story. Medium 3.5 collapses all of that. It is Mistral's first flagship merged model: a dense 128B model with a 256k context window, handling instruction-following, reasoning, and coding in a single set of weights, replacing both Mistral Medium 3.1 and Magistral in Le Chat, and Devstral 2 in the Vibe coding agent.
Reasoning effort is configurable per request, so the same model can answer a quick chat reply or work through a complex agentic run. You toggle between fast reply mode and deep reasoning mode per request , Mistral's version of extended thinking, where the model applies more test-time compute to harder problems when you ask it to. That per-request toggle is the key architectural convenience: no need to route between two models based on task complexity.
What it can do
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
