strix-mind

The intelligence is the box.

Run DeepSeek V4-Flash — 284 billion parameters, 1 million tokens of context — on the AMD Strix Halo desktop sitting under your desk. No cloud. No subscription. No data leaves your network. One install command.

▶ 60-second demo (video coming soon — shot list ready)

git clone https://github.com/Intuition-Labs-LLC/strix-mind
cd strix-mind && bin/mind-up

Open http://localhost:8080 in your browser. That's the chat.

---

Why this exists

You bought (or are considering) an AMD Strix Halo machine — a Framework Desktop, a Bosgame M5, a HP Z2 Mini G1a. 128 GB of unified memory. $3,400-ish. You wanted to run real AI on it, and the reviews you read couldn't agree on what was actually possible.

This repo is the answer. The model is DeepSeek V4-Flash (284B-parameter MoE, 1M native context, instruction-tuned, MIT-licensed, April 2026). The engine is bati.cpp with Vulkan kernels for the V4 architecture. The substrate is Mesa/RADV with RADV_DEBUG=novram — the one env-var setting that unlocks the unified memory pool. The agents are opencode, pi, hermes, and carl — best-in-class open-source surfaces, each pointed at one local endpoint.

We figured out what works. Now you don't have to.

What you get

• A private, multi-user web chat at http://localhost:8080 (the engine's built-in UI).
• DeepSeek V4-Flash loaded once, served forever, talking to four agent surfaces simultaneously through one OpenAI-compatible endpoint.
• Multi-agent concurrency via continuous batching — the engine never idles while there's work in the request queue.
• No metering. No telemetry. No subscriptions. Your box, your model, your data.

Surface	Purpose
opencode	Day-to-day coding TUI (150k★, 6.5M devs/month)
pi	Minimal agent loop; RPC dispatch target
hermes	Persistent personal agent — Telegram/Discord/Slack/Signal bridges, durable kanban
carl	Coherence-aware training studio — Phi telemetry on every generation

Performance, measured on a Framework Desktop

AMD Ryzen AI Max+ 395 / Radeon 8060S, 128 GB unified memory, Bazzite 42, kernel 6.17, Mesa 26.0.

Workload	Prefill	Decode
Single user, short prompt	—	~19 t/s
Single user, 322-token prompt	26.5 t/s	19.0 t/s
2 concurrent agents	—	9.6 t/s each (19.3 aggregate)
4 concurrent agents	—	3.3 t/s each (13.4 aggregate)

V4-Flash IQ2_XXS quantization (78 GB). f16 KV cache (required by V4's compressed-KV path). --batch-size 4096. Continuous batching enabled. Sweet spot: 1–2 concurrent agents.

Quick start

# 1. Clone
git clone https://github.com/Intuition-Labs-LLC/strix-mind
cd strix-mind

# 2. Install prerequisites (one-time)
#    See docs/INSTALL.md for the full sequence: bati.cpp build, model download,
#    agent-surface install. Roughly 30 minutes start to finish on a fast network.

# 3. Boot the appliance
bin/mind-up                 # idempotent; bring everything online
bin/mind-status             # green dots across the board
bin/mind-smoke              # one-shot prompt round-trips

# 4. Use it
#    Browser: http://localhost:8080  (built-in chat UI)
#    or via the agent surfaces:
opencode                    # coding TUI
pi -p "what's in pwd?"      # one-shot agent
hermes                      # personal-agent CLI
carl chat                   # Phi-aware conversation

Architecture

            user / household
                  │
        ┌─────────┼──────────┬──────────┬──────────┬─────────┐
        │         │          │          │          │         │
   Open WebUI   opencode    pi       hermes      carl    raw curl
   (browser)    (TUI)     (agent)   (personal)  (train)   (any)
        │         │          │          │          │         │
        └─────────┴────┬─────┴──────────┴──────────┴─────────┘
                       │  OpenAI /v1 over localhost:8080
                       ▼
           ┌───────────────────────────┐
           │  bati-server (V4-aware)   │  ← continuous batching, HTTP queue
           ├───────────────────────────┤
           │  DeepSeek V4-Flash IQ2_XXS│  ← 284B params, 1M ctx, 78 GB on disk
           ├───────────────────────────┤
           │  Vulkan / RADV STRIX_HALO │  ← RADV_DEBUG=novram for unified mem
           ├───────────────────────────┤
           │  AMD Ryzen AI Max+ 395    │  ← 16 cores, 128 GB unified, gfx1151
           └───────────────────────────┘

How it works (the substrate)

Three things that took us a while to figure out — and that this repo encodes so you don't have to:

1. RADV_DEBUG=novram collapses the small VRAM heap into the GTT pool. Without it, the Vulkan backend asks for domains: 4 (VRAM-only) allocations that don't fit, then OOM-cascades. With it, all 120 GiB of GTT is one unified pool that the engine can use.
2. Drop --mlock. Pinning every page of a 78 GB model on a 128 GB box leaves no headroom; mmap + the kernel page cache handle hot/cold pages efficiently because only ~13B parameters are active per token in V4-Flash's MoE.
3. f16 KV cache is mandatory for V4-Flash. Quantizing K/V to q4_0 or q8_0 produces multilingual token-salad output — V4's compressed KV path (ratio-4 CSA) has an f16 pin. Non-V4 models can safely use quantized KV; V4 cannot.

Full failure-mode catalogue with reproduction recipes: docs/TROUBLESHOOTING.md.

What strix-mind is NOT

• Not the inference engine. That's bati.cpp. Bug reports for the engine go upstream.
• Not the model. That's DeepSeek V4-Flash. Model behavior questions go upstream.
• Not a training framework. That's carl-studio, which strix-mind hosts.
• Not a multi-model router. One mind, loaded once, shared four ways. Swap models via bin/mind-swap-model or bin/mind-use-v4.

What strix-mind is: the orchestration layer that takes the upstream stack and makes it run as a single coherent appliance on AMD Strix Halo, with one command and a stable mental model.

Status

Component	State
V4-Flash IQ2_XXS load on bati Vulkan	✓ Working, validated
Substrate tuning baked into mind-up	✓ Working
Continuous batching across agent surfaces	✓ Working
One-shot smoke test (mind-smoke)	✓ Passing
Open WebUI integration (recipe)	✓ Documented, manual setup
Public install.sh wrapper	⏳ Planned for v0.2.0
antirez-style minimal C reference port	⏳ Planned for v1.0

See docs/STATUS.md for the live operational state and docs/TROUBLESHOOTING.md for everything we've personally broken.

Documentation

• docs/INSTALL.md — fresh-box install sequence, ~30 min
• docs/STATUS.md — current operational truth, benchmarks
• docs/TROUBLESHOOTING.md — every failure mode we've hit
• docs/ARCHITECTURE.md — engineering rationale
• docs/DEMO_SCRIPT.md — the 60-second video shot list
• NOTICE.md — credits with proper attribution

Credits

strix-mind is an orchestration layer over a tall stack of brilliant work by other people. Full attributions in NOTICE.md. The short version:

• bati.cpp — the only llama.cpp fork with working Vulkan V4-Flash kernels. Without bati, this appliance does not exist.
• llama.cpp (Georgi Gerganov + ggml authors) — the foundation. Continuous batching, the OpenAI-compatible server, the built-in chat UI we use at localhost:8080.
• DeepSeek AI — V4-Flash, MIT licensed, 284B parameters, the model.
• ssweens — the IQ2_XXS quantization (78 GB) that actually fits a 128 GB box.
• Mesa / RADV — the Vulkan driver. RADV_DEBUG=novram is theirs.
• opencode, pi, hermes-agent, carl-studio — the four agent surfaces.
• antirez — for proving over decades that small, hackable, sincerely-cited code is the right shape for systems software. We try to follow the form.

License

MIT. © 2026 Intuition Labs LLC. See LICENSE.

Compatible with every upstream we depend on. Take this, fork it, ship something better.

---

Built by Tej Desai with AI collaboration for Intuition Labs LLC. The intelligence is the box.