# Run DeepSeek-V4-Flash on your own computer, no GPU needed

> llama-modchip runs DeepSeek's 284B open model on 64 GB of RAM with no graphics card, plus Zyphra's shared team agent MAIA, Slack and Teams cards for Chat SDK bots, and a cleaner way to review AI-written pull requests.

- Published: 2026-11-09
- Publisher: Rig Labs (https://riglabs.ai/)
- Canonical: https://riglabs.ai/briefs/2026-11-09/

---

## llama-modchip

**Run DeepSeek-V4-Flash, a 284B-parameter open model, on a computer with 64 GB of RAM and no graphics card.**

**Why:** your prompts stay on your machine, and you pay no per-message fee.

**How:** llama-modchip, a free open-source tool by synchopate, squeezes the model into a file of about 61 GB. Its author reports about 9 to 15 tokens a second on an AMD Ryzen 9 9955HX processor.

**Catch:** it only runs with the project's own software. Ollama and LM Studio cannot open these files. The speed figures come from the author and community reports.

<details><summary>Technical detail</summary>

Version 2.4.0, 2026-11-04, Apache 2.0, built on llama.cpp. The model has 13B active parameters and is stored at 1.71 bits per weight (56.56 GiB), with a measured peak of 57.61 GiB of memory at 256K context, on an AMD Ryzen 9 9955HX. About 44 tokens a second when reading a prompt. A smaller build of Nemotron-3.5-30B-A3B needs 14.35 GiB. Discussed on [Hacker News](https://news.ycombinator.com/item?id=50411961) on 2026-11-08.

</details>

[The v2.4.0 release](https://github.com/synchopate/llama-modchip/releases/tag/v2.4.0), 2026-11-04. [The repo](https://github.com/synchopate/llama-modchip).

Our bet: more people will try a large open model on the desktop they already own before they rent a GPU.

## MAIA by Zyphra

**Give your whole team one shared AI agent in Slack that remembers how your company works.**

**Why:** everyone works with the same agent, so it learns your team's decisions and ways of working once, in what Zyphra calls Company Memory.

**How:** Zyphra says MAIA connects to the tools a team already uses and acts in them, builds small apps the team can open and change, and asks before any action that cannot be undone. It is in beta and free during the beta.

**Catch:** the page names MAIA for Slack and no other integration yet.

[Zyphra's page](https://zyphra.com/maia), announced 2026-11-06.

## json-render for Chat SDK

**If you build a Slack or Teams bot, it can now answer with real tables, charts, fields and buttons instead of plain text.**

**Why:** people can scan and click the answer, and the cards fill in while the AI is still writing.

**How:** Vercel connected json-render, its library for generating interfaces, to Chat SDK agents with the `@json-render/chat` package. The agent writes a JSON spec, which becomes Block Kit on Slack and Adaptive Cards on Teams. If the spec breaks the rules, `toCard` returns a readable list of problems the model can fix.

[The changelog](https://vercel.com/changelog/json-render-and-chat-sdk-agents), Vercel, 2026-11-06.

## A habit from Simon Willison

**Make AI-written pull requests easy to review: your own two-line summary on top, the agent's long write-up folded below.**

**Why:** Simon Willison now gets most of his changes as pull requests opened by Claude Code or Codex. A reviewer reads his short summary first, sometimes with a screenshot, and opens the agent's detail only if needed.

**How:** he keeps the agent's full write-up (problem, solution, how it was checked) inside a block that GitHub shows collapsed, labelled with the model name:

```
<details><summary>Written by Claude Code</summary>

(the agent's problem, solution and checks go here)

</details>
```

[Simon Willison's post](https://simonwillison.net/2026/Nov/9/details-summary/), 2026-11-09.
