Claude Code, but the engine is silicon-baked Llama 3.1 8B doing ~14,500 tokens/sec.
A tiny local proxy that speaks the Anthropic Messages API, so the full Claude Code harness
drives the fastest model on Earth. Half lobster, half Lamborghini, all throughput.
Website · Quickstart · How fast · How it works · The story · Safety
Taalas bakes AI models directly into custom silicon — "the model is the computer," ~1000× more efficient than running it on GPUs. Their public demo (which the world knows as ChatJimmy) serves Llama 3.1 8B etched into a chip at an almost silly speed: ~14,500 tokens/sec, ~1.1 ms to first token.
claudberghini straps that engine to Claude Code. It's a
local proxy that:
- Speaks the Anthropic Messages API — so Claude Code (via deep-claude) talks to it unmodified.
- Translates tool calls — injects a tool-use format into the prompt and parses the model's text back into Anthropic
tool_useblocks (the demo has no tools API). - Makes a weak 8B reliable — best-of-N resampling, grounded answer selection, a tuned compact prompt, and a destructive-command guard.
The result is a real agentic coding loop — read, edit, create, search, run — on a model so fast that drawing it 5× per turn for reliability is still faster than a single Claude Opus call.
Same task — "write a basic HTML page" — measured end to end:
| Engine | Tokens/sec | vs claudberghini |
|---|---|---|
| claudberghini (Taalas silicon, Llama 3.1 8B) | ~14,500 | — |
| Groq (Llama 3.1 8B) | ~1,250 | 12× slower |
| Claude Sonnet 4.5 | ~50 | ~290× slower |
| Claude Opus 4.1 | ~14 | ~1,000× slower |
A full HTML page is ~8 ms of inference on claudberghini. Claude Opus spends ~9 seconds on the same tokens. That's the headline: ~1,000× faster than Opus — which, fittingly, is exactly the order of magnitude Taalas claims for silicon-baked models.
Reproduce: ./eval/speed-benchmark.sh (claudberghini telemetry + end-to-end); the Opus/Sonnet
numbers via OpenRouter.
- macOS or Linux · Claude Code on your
PATH· deep-claude · Node.js 18+
git clone https://github.com/dennisonbertram/claudberghini
cd claudberghini
npm install && npm run build
# register the endpoint with deep-claude (one time)
deep-claude endpoints add claudberghini http://localhost:3000cd your-project # run it from the project you want to work in
/path/to/claudberghini/claudberghiniThat's it. The claudberghini launcher auto-starts the proxy and opens a clean Claude Code
session on the silicon. One-shot mode:
claudberghini -p "create an index.html with a Hello World heading"
# …done in ~1.5s.Alias it:
echo 'alias cb="/path/to/claudberghini/claudberghini"' >> ~/.zshrcRun it from your project directory. Claude Code sandboxes file tools to the working directory — paths outside it (like
/tmp) are blocked.
Claude Code (Anthropic Messages API)
│ via deep-claude --endpoint claudberghini (isolates your real Anthropic login)
▼
claudberghini proxy :3000
• swaps Claude Code's 120 KB prompt for a compact, eval-tuned one
• strips <system-reminder> noise, filters ~60 tools down to the coding set
• injects a <tool_call> format into the system prompt
• best-of-N: re-sample until a valid tool call parses
• grounded best-of-N: pick the answer most supported by tool output
• guards against destructive commands
▼
Taalas silicon (the "ChatJimmy" demo) → Llama 3.1 8B baked into a chip @ ~14,500 tok/s
Backend-switchable: BACKEND=openrouter routes to meta-llama/llama-3.1-8b-instruct instead —
the same model on a billed API, which is how we tuned without hammering the demo.
We couldn't get an API key. Taalas's demo doesn't ship one — it's a chat box on a website. So we
did the only reasonable thing: opened the browser dev tools, watched it talk, and rebuilt the
API from the network traffic. (POST /api/chat, raw-text streaming, a <|stats|> trailer with
the juicy token-rate telemetry. You're welcome.) We didn't get access; we made access.
Then the hard part. An 8B model is a mediocre tool-follower out of the box — it rambles, invents
filenames, botches JSON, and occasionally tries to sudo rm your /etc/hosts when you say "hi."
So we built an eval harness (real Claude Code agent loops over verifiable coding tasks) and
iteratively, recursively tuned against it:
- a hill-climbing workflow proposed prompt variants, scored each on the eval, kept the winners, and looped until 10 rounds passed with no improvement;
- we trimmed the toolset the model sees from ~60 down to the coding essentials (the noise was making it pick the wrong tool);
- we taught the proxy to re-sample on a botched tool call and to ground final answers in real tool output (killing the "I'll just make up the filename" failure mode);
- and we added a guard so the fast-but-dim model can't be talked into anything destructive.
Net result: the core operations — read, edit, create, grep — land 5/5, on the fastest inference on the planet. The eval/ directory has the whole rig.
| Task set | claudberghini (Taalas) | OpenRouter llama-3.1-8b |
|---|---|---|
| Core 4 (read / edit / create / grep) | 5/5 each | 5/5 |
| Hard 10 (multi-step, mutations, multi-file) | ≈0.70 | 1.00 |
It's an 8B model: superb for focused file ops, weaker on tangled multi-file logic. The Taalas demo instance is more quantized than OpenRouter's, so absolute quality is a touch lower — but the core loop is reliable, and nothing else is remotely this fast.
A fast, weak model can hallucinate commands, so the proxy refuses to pass through destructive or
privileged shell calls (sudo, rm -rf /, mkfs, dd, curl … | sh, writes to
/etc/{passwd,shadow,hosts,sudoers}). It also never forces a tool call — "hi" gets a plain
reply, not an invented command. Claude Code's own permission system still applies on top.
Defense-in-depth for a small model, not a security boundary — run it on code you trust.
| Env | Default | What |
|---|---|---|
BACKEND |
claudberghini |
openrouter to use meta-llama/llama-3.1-8b-instruct |
TOOL_SAMPLE_ATTEMPTS |
5 |
best-of-N draws to get a valid tool call |
ANSWER_SAMPLE_ATTEMPTS |
3 |
grounded answer candidates |
TOOL_ALLOWLIST |
coding set | comma list, or * to pass all tools through |
MAX_SYSTEM_BYTES |
18000 |
trim the system prompt to fit the ~24 KB ceiling |
CLAUDBERGHINI_API_URL |
https://chatjimmy.ai |
the Taalas demo endpoint |
Inference by Taalas — the model, baked into silicon. Built to ride on deep-claude, which points Claude Code at a custom endpoint while keeping your real Anthropic login untouched.
MIT — see LICENSE.