paritok 1.2.6

Open-source context compression for AI coding agents. A middle layer between your agent and the LLM API — self-host the 4B model or use the Paritok GPU server.

Paritok

The first open-source compression model trained specifically for coding agents.

Trained on 45K real coding-agent trajectories, Paritok understands the difference between a function signature and a debug line — so it keeps what matters and drops what doesn't. Fully compatible with Claude Code, Cursor, OpenHands, and any agent framework using standard message format.

~74% fewer tokens on typical workloads (up to 95% on heavy long-session traffic), cutting your input token bill by up to 95% on Claude / GPT — while matching gpt-4.1-mini on SWE-bench Verified at a fraction of the cost.

Highlights · Benchmark · Cost · Compare · Quick Start · Model Card · Training · Team · Citation

---

📢 News

• 2026-07-14   Paritok-4B-v1 released on Hugging Face Hub with full SWE-bench Verified end-to-end evaluation.
• 2026-06-25   Finished training. 45K teacher-distilled samples on the Qwen3-4B backbone.

---

✨ Highlights

• 🎨 Code-native. Trained end-to-end on real coding-agent trajectories (file_read, bash_command, log_output, ...). Paritok knows what an import statement is worth vs a debug line, so it protects function names, paths, and error strings while compressing.
• 🚀 Up to 95% fewer tokens (74% on typical workloads) — compresses each segment to 25.7% of original; drop-aware long sessions push overall cuts to 95%. 2× harder than gpt-4.1-mini (50.2% CR) and 2.4× harder than gpt-5 (61.9% CR).
• 🎯 Retains 86.5% of full-context solve quality on SWE-bench Verified — matching gpt-4.1-mini as compressor at less than half the token spend, and within 7pp of gpt-5 at 40% its context length.
• 💰 Up to 95% off your input token bill (74% on typical workloads) at Claude Sonnet pricing. Long-session teams save thousands per month — see Cost Impact.
• 🪶 Small & self-hostable — 4B LoRA adapter, bf16, runs on a single 24GB GPU. No SaaS, no lock-in, no per-token compressor fee.
• 🔓 Fully open — Apache 2.0 weights, reproducible data pipeline, real end-to-end SWE-bench numbers (no cherry-picking).

---

📊 Benchmark: SWE-bench Verified

Real end-to-end evaluation on SWE-bench Verified. An agent scaffold receives its context through each compressor, then attempts to resolve the issue. Primary metric is quality retained (solve rate normalized to the uncompressed baseline) — the fair way to compare compressors of different aggressiveness.

Solve quality vs compression rate

Context source	Quality retained ¹	Compression rate
Uncompressed baseline	100.0%	100.0%
gpt-4.1-mini (compressor)	85.6%	50.2%
gpt-5 (compressor)	93.6%	61.9%
Paritok-4B-v1 ⭐	86.5%	25.7%

_{¹ Quality retained = compressor solve rate ÷ uncompressed baseline solve rate. Higher is better.}

What this says:

• Paritok retains 86.5% of uncompressed solve quality — matching gpt-4.1-mini's retention — while shipping less than half the tokens.
• Paritok comes within 7pp of gpt-5's retention at ~40% of gpt-5's context length.
• No open compressor comes close on the CR / quality frontier.

Paritok cuts context by 74% while retaining 86.5% of uncompressed solve quality — enough headroom to double or triple your monthly turn budget on the same API spend.

_{Benchmark numbers above use Paritok's baseline 25.7% CR (74% context cut) — the raw compressor performance in a controlled setting. Real production savings can go higher with drop-aware deployment; see Cost Impact for the 95% upper bound.}

Evaluation used the standard SWE-bench Verified harness with a public agent scaffold. Per-issue results and reproduction instructions will be published with the v2 release.

---

💰 Cost Impact

Cut your input token bill by up to 95%. Paritok compresses input; output is unchanged and left to your provider.

How to read the numbers below: all tables use Paritok's typical 74% saving (median compression rate 25.7%). Long agent sessions with heavy tool-output and file-read repetition push savings up to 95% — real bills often land higher than shown here.

At Claude Sonnet input pricing ($3 / M input tokens), typical scenario (74% saving):

Turn input size	Uncompressed input cost	With Paritok
Short (8K)	$0.024	$0.006
Typical (15K)	$0.045	$0.012
Long session (30K)	$0.090	$0.023

Real project scenarios (typical 15K/turn workload, 74% saving)

Scenario	Uncompressed input	With Paritok	Saved
Solo dev, 1-week prototype (5d × 300 turns)	$67.50	$17.34	$50
Startup, 1-month project (20d × 400 turns)	$360	$92	$268
10-person team, 3-month project (60d × 10 × 500)	$13,500	$3,468	~$10K

Deployment overhead pays for itself in days, not weeks — and there's no lock-in: it's your own 4B model on your own hardware.

---

🆚 How Paritok compares

	Paritok-4B-v1	LLMLingua-2	gpt-4.1-mini prompt
Trained on real coding-agent trajectories	✅	❌	❌
Preserves function names / imports / paths	✅ (by design)	partial	partial
Compression rate (lower = harder)	25.7% ⭐	~40%	50.2%
SWE-bench Verified — quality retained	86.5% ⭐	not evaluated	85.6%
Self-hostable open weights	Apache 2.0	MIT	closed API
Per-token compressor fee	zero (self-host)	zero (open)	pay-per-token

Paritok is the only entry trained end-to-end on real coding-agent trajectories — that's why it compresses ~2× harder than gpt-4.1-mini prompting on the same task while keeping the same solve rate.

---

🔧 Cache-friendly tool selection (embedding)

Coding agents (Claude Code, Cursor, …) expose dozens of tools — often 70+ once you add MCP servers — on every request. That tool-schema JSON is a large slice of input tokens that's mostly irrelevant to the task at hand. Paritok can filter it semantically: set tool_discovery.strategy: embedding and it keeps only the handful of tools relevant to the user's intent in full schema, stubbing the rest.

Crucially it's prompt-cache friendly — the selection is frozen per conversation, so the tools[] block stays stable turn-to-turn and doesn't invalidate the LLM's KV cache. Anything filtered out is still recoverable on demand (the model calls gateway_search_tools).

It runs a small open embedding model, BAAI/bge-small-en-v1.5 (MIT-licensed, ~130MB), entirely locally on CPU — no API, no per-token fee.

This is the default tool-discovery strategy (paritok init writes it for you):

tool_discovery:
  strategy: embedding

pip install "paritok[toolselect]"   # adds sentence-transformers (CPU-only)

First request warms up the model (~10–15s, depending on network speed — it downloads bge-small from Hugging Face once, then caches locally). Every request after that is instant (~15 ms).

---

🚀 Quick Start

Paritok runs as a middle layer between your agent and the LLM API. It intercepts each request, compresses the context, and forwards it upstream — your agent doesn't change, it just points at Paritok.

Your Agent (Claude Code / Cursor / Codex)
  → builds request (tool results + history + tool schemas)
     ★ Paritok middleware compresses here ★
  → forwarded to Anthropic / OpenAI  (billed on the compressed tokens)
  ← response flows back unchanged; compressed refs expand on demand

Fastest path (self-host, no clone needed)

Everything ships in the PyPI package — you do not need to git clone the repo. In a fresh environment (with Ollama installed):

pip install "paritok[proxy]"     # the middleware + CLI
paritok up                       # pulls the model if missing, then starts the proxy

paritok up is the pip-only, no-clone equivalent of a setup script: it checks Ollama, ollama pulls paritok/paritok-4b-v1 and tags it as the local paritok-4b-v1 if it isn't already there (~2.5GB, first run only), then serves on port 8080.

Leave that terminal running. The proxy is a foreground server — every agent request flows through it, so it must stay up for the whole session. Closing the terminal (or Ctrl-C) stops compression. Open a separate terminal for the next step.

In the shell that launches your agent (step 4):

export ANTHROPIC_BASE_URL=http://127.0.0.1:8080   # then start Claude Code / Cursor / …

No config file is needed — paritok up and paritok proxy run on built-in defaults. Run paritok init to drop a starter paritok.yaml only if you want to tweak settings. (Cloned the repo? ./deploy.sh does the same pull + start.)

q4 vs full precision. paritok up uses the q4 model (:latest, ~2.5GB) by default. For full precision, run paritok up --registry-model paritok/paritok-4b-v1:f16 (~8GB). If you already ollama pulled either variant yourself, up auto-detects which one you have and uses it — no re-download. Prefer to run each step by hand? They're below.

1. Install

pip install "paritok[proxy]"
# or, from a clone of this repo:
pip install -e ".[proxy]"

2. Pick a backend — self-host or the GPU server

One boolean in paritok.yaml decides where compression runs:

use_gpu_server: false   # ← the only switch that matters

Option A — self-host (false, default). Run the open 4B model on your own machine. No key, nothing leaves your box. Simplest is Ollama:

ollama pull paritok/paritok-4b-v1               # one-time, ~2.5GB
ollama cp   paritok/paritok-4b-v1 paritok-4b-v1  # tag it as the runtime name

The registry (pull) name is namespaced; the second line tags it as the bare paritok-4b-v1 the config uses. Default paritok.yaml already points local_model at Ollama (http://localhost:11434/v1, model: paritok-4b-v1) — nothing else to set. (./deploy.sh does both steps for you.)

Want full precision? Pull paritok/paritok-4b-v1:f16 instead (~8GB, needs more RAM/VRAM). The default :latest is q4_K_M (~2.5GB) — the quantization the runtime is validated against.

Alternative: vLLM (serves the HF LoRA adapter directly, no GGUF)

The Hugging Face weights (paritok/paritok-4b-v1) are a LoRA adapter over Qwen/Qwen3-4B-Instruct-2507. vLLM can serve it as an OpenAI-compatible endpoint on a 24GB GPU:

pip install vllm
vllm serve Qwen/Qwen3-4B-Instruct-2507 \
  --enable-lora \
  --lora-modules paritok-4b-v1=paritok/paritok-4b-v1 \
  --max-lora-rank 32 \
  --port 8000

Then in paritok.yaml, set local_model.base_url: http://localhost:8000/v1 (keep model: paritok-4b-v1, the --lora-modules name).

Whatever you name the served model, make local_model.model match it (or override once with PARITOK_MODEL=<name>).

Option B — Paritok GPU server (true). No GPU required. Create an API key at paritok.com → dashboard → API keys, then:

use_gpu_server: true
gpu_server:
  api_key: "pk_live_..."   # or: export PARITOK_API_KEY=pk_live_...

3. Start the proxy

paritok proxy --port 8080 --config-file paritok.yaml

Keep this terminal open — the proxy must stay running for the whole session; run your agent from a separate terminal. On startup it checks the backend and warns (never aborts) if it can't reach one — e.g. the Ollama model isn't pulled, or the hosted server isn't reachable.

4. Point your agent at it

Set the base URL in the shell that launches your agent, then start the agent:

# macOS / Linux
export ANTHROPIC_BASE_URL=http://127.0.0.1:8080   # Claude Code
export OPENAI_BASE_URL=http://127.0.0.1:8080      # Cursor / OpenAI-SDK agents

# Windows PowerShell
$env:ANTHROPIC_BASE_URL = "http://127.0.0.1:8080"
$env:OPENAI_BASE_URL    = "http://127.0.0.1:8080"

Keep your real provider API key set as usual (ANTHROPIC_API_KEY / OPENAI_API_KEY) — the proxy only rewrites the request body and forwards your headers upstream. That's it: Claude Code, Cursor, and any agent that honors BASE_URL now routes through Paritok. Compressed prompts go upstream; original responses come back unchanged.

Codex is the exception — it does not read OPENAI_BASE_URL. The Codex CLI takes its endpoint from ~/.codex/config.toml, not an env var, so the line above silently has no effect for it. See Codex setup just below.

Any OpenAI-compatible upstream (Groq, Gemini, …)

The compressor is independent of the upstream LLM — it runs locally on the Paritok 4B model (or the GPU server), so you can point the OpenAI path at any OpenAI-Chat-Completions-compatible endpoint with --openai-url. Your agent sends its provider key as usual; the proxy forwards it.

# Groq — base host; the proxy appends /v1/chat/completions
paritok proxy --openai-url https://api.groq.com/openai
#   -> https://api.groq.com/openai/v1/chat/completions

# Gemini — its OpenAI-compat path isn't {base}/v1/..., so pass the FULL endpoint
paritok proxy --openai-url https://generativelanguage.googleapis.com/v1beta/openai/chat/completions

--openai-url accepts either a base host (standard /v1/chat/completions is appended) or a full endpoint already ending in /chat/completions (used verbatim). Then point your OpenAI-SDK agent at the proxy (OPENAI_BASE_URL=http://127.0.0.1:8080), set OPENAI_API_KEY to the provider's key, and use that provider's model names.

Codex CLI

Codex ignores OPENAI_BASE_URL — it only reads its endpoint from ~/.codex/config.toml. So Paritok writes that file for you. Flip one switch in paritok.yaml and paste your key:

codex:
  enabled: true            # paritok writes ~/.codex/config.toml on `paritok up`
  model: gpt-5             # any model your key can call
  api_key: "sk-..."        # your OpenAI key (or leave "" to use env OPENAI_API_KEY)

Then start the proxy and run Codex — that's it:

paritok up     # writes ~/.codex/config.toml (backs up any existing one)
codex          # in another shell — now routed through Paritok

Everything lives in paritok.yaml; the generated config.toml is a derived file (a .paritok-bak backup is kept if you already had one). Codex custom providers only speak the responses wire protocol, which the proxy serves at /v1/responses.

Prefer to write ~/.codex/config.toml by hand?

model = "gpt-5"
model_provider = "paritok"

[model_providers.paritok]
name = "paritok"
base_url = "http://127.0.0.1:8080/v1"
wire_api = "responses"
env_key = "OPENAI_API_KEY"      # your OpenAI key, read from the environment

Using an API key instead of a subscription

Signed in with a subscription (Claude Pro/Max, or ChatGPT via codex login)? Just launch claude / codex the way you normally do.

Prefer a pay-as-you-go API key?

• Claude Code — set your key and keep the base URL from step 4:

  export ANTHROPIC_API_KEY=sk-ant-...
  export ANTHROPIC_BASE_URL=http://127.0.0.1:8080
  claude
  

• Codex — put your key in the codex: block of paritok.yaml (above); nothing else to set.

5. Check it's working

The proxy exposes two endpoints:

curl http://127.0.0.1:8080/health   # {"status":"ok","version":"..."}
curl http://127.0.0.1:8080/stats    # live compression totals

/stats reports cumulative savings across the session:

{
  "total_requests": 42,
  "input_tokens_original": 512340,
  "input_tokens_compressed": 138221,
  "compression_ratio": 0.27,
  "tokens_saved": 374119,
  "estimated_cost_saved_usd": "$1.01"
}

These numbers are scoped to what Paritok actually intervenes in — the content it compresses (tool results / file reads / old history) plus the tool schemas it stubs. Everything it can't affect (your system prompt, the model's output) is deliberately excluded. compression_ratio and tokens_saved are raw token counts on that domain (lower ratio is better).

estimated_cost_saved_usd is cache-aware, not a flat list-price multiply, priced at each model's own input rate (the proxy sees the model on every request; unknown → $3/M):

• Compressed content (file reads / tool output / history) is new input each turn, so its saving is priced at the full base rate.
• Tool schemas are frozen per conversation (the selection is stable byte-for-byte across turns), so that block is a prompt-cache hit: its first turn is a cache write (1.25× base) and every turn after is a cache read (Claude 0.1×, GPT-5 0.1×, gpt-4o 0.5×, …). Counting it at full list price would massively overstate savings, since the uncompressed schemas would have been cached too.

Edit the base rates and cache multipliers in paritok/proxy/pricing.py. The hosted dashboard at paritok.com reports the same content + tool basis for use_gpu_server: true traffic.

SDK mode (alternative)

Prefer to wrap your client directly in Python?

import anthropic
import paritok

client = paritok.ParitokClient(anthropic.Anthropic())
resp = client.messages.create(
    model="claude-sonnet-4-20250514", max_tokens=4096, messages=[...]
)
print(resp._paritok_savings.saved_tokens, resp._paritok_savings.ratio)

Just the raw model?

To load the LoRA adapter and compress a single [SEG] block yourself (no middleware), see examples/inference/basic.py.

---

🧩 Use Cases

Paritok is most useful when:

• Your AI coding agent (Claude Code / Cursor / Copilot / OpenHands / custom SDK) sends > 5 000 tokens per turn.
• You're paying per token to Anthropic / OpenAI / other API providers.
• You want lower per-turn latency (fewer input tokens = faster prefill).
• You can tolerate ~300 ms of compression overhead per request.

Paritok is less useful when:

• Your context is already short (< 2 000 tokens).
• You need lossless compression (in that case, just don't compress).
• Your workflow is single-turn Q&A (context doesn't accumulate).

How the middleware works

The paritok/ package is the middle layer. On every request the engine (paritok/middleware/wrapper.py) runs four steps before forwarding upstream:

1. Tool discovery — 70+ tool schemas → top-K kept in full, the rest stubbed (pipelines/tool_discovery.py).
2. Compress tool outputs — each tool_result is compressed by the 4B model (pipelines/compress.py).
3. Compress old history — turns beyond the recent window are summarized once the context fills up.
4. Inject virtual tools — expand_context and gateway_search_tools (pipelines/virtual.py) let the model pull back anything it needs.

Never destructive. Compressed content is tagged [REF:id]; if the LLM needs the original it calls the expand_context virtual tool and the middleware returns the full text locally. gateway_search_tools recovers any tool schema that discovery stubbed out.

Deploy it either way — the same package powers both:

Your Agent  ──►  Paritok middleware  ──►  Anthropic / OpenAI
   (raw)          (self-host or GPU server)   (billed on compressed)

Configure everything in paritok.yaml; flip use_gpu_server to move between your own hardware and our hosted GPU endpoint without touching your agent.

---

📋 Model Card

Property	Value
Base model	Qwen/Qwen3-4B-Instruct-2507
Adapter type	LoRA, r=32, α=64, dropout=0.0
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training steps	2000 (selected from a 5-checkpoint sweep, best on SWE-bench Verified subset)
Training precision	bf16
Effective batch	32 (per_device=2 × grad_accum=16)
Learning rate	1e-5, linear decay, 10% warmup
Optimizer	AdamW (8-bit)
Max seq length	16 384
Dataset size	45 000 samples across file_read, bash_command, log_output, etc.
Teacher	gpt-4.1-mini
Weights	🤗 HF Hub
License	Apache 2.0 (weights); base model under its own Qwen license

Full training config: training/configs/sft_config_qwen3_4b.yaml.

---

🎓 Training

Reproduce from scratch

High-level pipeline; see training/ and data_pipeline/ for the actual scripts.

# 1. Prepare data (regenerate pools from agent-trajectory dumps)
python data_pipeline/extract/extract_file_read_pool.py --n 10000
python data_pipeline/extract/extract_other_kinds_pool.py

# 2. Distill via teacher (requires OPENAI_API_KEY, ~$300 in API cost)
python data_pipeline/compress/compress_pool_file_read.py
python data_pipeline/compress/compress_pool_other.py

# 3. Train SFT (2× A100 80GB or 1× H100 80GB, ~5 hours)
bash deploy_sft_4b.sh

Or on a fresh RunPod / Lambda / Modal pod:

RUNPOD_HOST=<host> RUNPOD_PORT=<port> ./deploy_sft_4b.sh

Pipeline summary

1. Data collection — 100k+ raw agent-trajectory turns from open-source SWE-bench-style trajectory dumps.
2. Segmentation — split each turn into [SEG] blocks by kind (file_read, bash_command, log_output, ...).
3. Teacher distillation — gpt-4.1-mini produces the target compression for each segment.
4. Filter & rebalance — drop mal-formatted teacher outputs, up-sample rare kinds.
5. SFT — LoRA on Qwen3-4B-Instruct-2507, imitation loss on the teacher's compressed reply.
6. Checkpoint selection — evaluate on SWE-bench Verified end-to-end, pick step 2000.

---

🗺️ Roadmap

Paritok-4B-v1 is our first release. What's next:

• 🎯 Paritok-4B-v2. Next-generation training pipeline pushing compression to under 20% while closing the gap to uncompressed solve rate. Target: up to +15pp identifier retention at even tighter compression.
• 📈 Frontier-scale backbones. Larger models (10B+ parameters) for multi-day Claude Code / Cursor sessions with 100K+ token histories and heavy multi-file workflows.
• 🌍 Multi-language expansion. First-class support for TypeScript, Rust, Go, Java, C++, Kotlin — v1 is Python-heavy but the architecture is language-agnostic.
• 🔌 Native integrations. Drop-in mcp add paritok plugin for Claude Code and Cursor. (Or route to the hosted GPU endpoint with use_gpu_server: true.)
• ⚙️ Adaptive compression. Per-segment auto-selection of compression aggressiveness based on age, kind, and downstream intent — no manual tuning, no level knobs.

Follow the 🤗 model discussions or star the repo for release notifications.

---

👥 Team

Paritok is built by two engineers — no big lab, no external funding, just months of GPU budget and eval iteration.

• Jiayu Shi — training, modeling, reward design, data pipeline.
• Luzhuo Chen — evaluation, deployment, product, data pipeline.

We ship on our own budget and share every result transparently. Paritok-4B-v1 is our first release; v2 is in training.

Reach us: [e-mail] · X @Paritok

---

📖 Citation

If you find this work useful, please cite:

@misc{paritok-4b-v1,
  author       = {Paritok Team},
  title        = {Paritok-4B-v1: An Open-Source Compression Model for AI Coding Agents},
  year         = {2026},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/Paritok-official/paritok-4b-v1}},
}

Please also cite the base model:

@misc{qwen3-4b-instruct,
  author       = {Qwen Team},
  title        = {Qwen3-4B-Instruct},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507}},
}

---

📄 License

Apache 2.0 — see LICENSE.

The base model, Qwen3-4B-Instruct-2507, is released under its own license. Please review it before commercial deployment.

---

💬 Community & Support

• 🐛 Bug reports & feature requests → GitHub Issues
• 💭 Discussion → 🤗 HF Model discussions
• 📧 Contact → [e-mail]