A Llama-family reader handed usable memory to a Qwen-family receiver.

We asked TinyLlama-1.1B to read one of eight synthetic records, translated its key-value cache, and gave only that translated state plus the question to Qwen3-4B-Instruct. Across 192 fresh held-out questions, the transferred path answered 187 correctly. Native Qwen prefill answered 192.

The 2.6-point gap cleared the predeclared five-point acceptance limit. All twelve frozen gates passed, including source controls, cache geometry, receiver forward traces, and counterfactual tests.

Native Qwen3-4B192 / 192
Cross-family transfer187 / 192
No context18 / 192
Zero cache0 / 192

The translator crossed every obvious compatibility boundary.

This is not a handoff between checkpoints from the same family. TinyLlama and Qwen3 use different tokenizers, layer counts, KV-head counts, head dimensions, and rotary-position configurations. Both models stayed frozen. The only learned component was an 82,649,088-parameter cache translator.

Reader TinyLlama 1.1B SentencePiece · Llama cache geometry
82.6M learned translator
Receiver Qwen3 4B Qwen tokenizer · Qwen3 cache geometry

The translator removes source rotary positions, aligns text spans across tokenizations, maps layer and channel geometry, and reapplies target rotary positions. At inference, it uses source keys and values only. Native target prefill is supervision for training and a control for evaluation, not an input to the transferred path.

Changing the cache changed the answer.

High accuracy is not enough to prove that the receiver used transferred memory. We replaced the source record with a different record that assigns a different answer to the same question. The receiver followed the swapped fact in 191 of 192 cases and repeated the original answer in zero.

Fresh 192-case held-out evaluation
ArmExpected behaviorObserved
Mapped source cacheAnswer from the original record187/192
Swapped source cacheFollow the counterfactual record191/192
No contextControl floor18/192
Zero cacheControl floor0/192
What this establishes

The translated cache carried task-relevant facts across model families. It did not merely perturb Qwen toward a memorized answer distribution.

Qwen answered in a separate process that never saw the record.

We serialized a 13.6 MB transferred cache, then launched a separate Qwen-only receiver. That process received no TinyLlama model and no original context text. It imported the cache, answered the excluded warmup question correctly, and produced bit-identical first-token logits to the in-process path, with maximum difference zero.

The imported cache represented a 92-token prefix. Qwen's recorded forward lengths were [15, 1]: fifteen suffix/question tokens, then one token at a time during generation. The receiver never replayed the 92-token context.

  1. 01Read

    TinyLlama processes the record once.

  2. 02Translate

    The adapter writes a portable Qwen cache.

  3. 03Detach

    A fresh process loads Qwen only.

  4. 04Answer

    Qwen reads the question, not the record.

This experiment proves communication, not acceleration.

Median end-to-end time to first token was 278.81 ms for the transferred path and 258.33 ms for native Qwen prefill on this short-context M2 Max pilot. The paired native-to-transfer ratio was 0.923, so the transfer was slower.

That negative result is part of the finding. A portable cross-family memory now works at high accuracy. Making it useful for serving requires longer contexts, a smaller translator, and an implementation where avoided target prefill outweighs source prefill plus mapping.

The claim boundary

This is an eight-record synthetic-recall result between two decoder Transformers. It is not evidence for arbitrary language transfer, larger receivers, or lower production latency.

Move from a memory primitive to a useful inference path.

  • Test natural documents and continuations where exact-string recall is not enough.
  • Scale the receiver from Qwen3-4B to Qwen3-14B under a fresh frozen protocol.
  • Measure the context-length crossover where target-prefill savings exceed translation cost.
  • Compress the 82.6M-parameter translator without weakening the counterfactual controls.
  • Preserve the receiver-only process boundary as the acceptance test for every scale-up.