Evidence for deploying compressed, governed AI.

Showing all 9 write-ups, newest first.

1. R-012
   28 October 2026 · Benchmark

   We shrank GLM-5.3-Flash by 68%, and it solved 86% of SWE-bench Verified.

   GLM-5.3-Flash is 642.65 GB in 16 bits; Tileward’s TWFP4 build is 204.45 GB and runs on two B300 cards. As a coding agent it solved 431 of 500 SWE-bench Verified tasks, and 429 on a repeat. The original was not run on SWE-bench; TWFP4 picks its next word 88.8% of the time.

   Agent workflowsModel compressionQuality
   Experiment

   204.45 GB

   of weights, against 642.65 GB in 16 bits

   431 of 500

   SWE-bench Verified tasks solved

   429 of 500

   solved when the run was repeated

   Take a look →
2. R-011
   28 October 2026 · Benchmark

   We shrank Kimi K3 by 35%, within a point of the original on MMLU.

   Kimi K3 is 1,560.9 GB as Moonshot publishes it, already stored in 4 bits. Tileward’s TWX2.5 build is 1,007.9 GB, 90.5% on MMLU against 91.1%: 0.6 points below, inside the one-point bar set before the run, and level on made-up facts and tool calls. Our TWX3 build keeps closer to the original.

   Model compressionMMLUQuality
   Experiment

   1,007.9 GB

   of weights, against 1,560.9 GB as published

   90.5%

   on MMLU, against 91.1% for the original

   2.7×

   the original’s cache, on the same eight cards

   Take a look →
3. R-010
   28 October 2026 · Benchmark

   We made Qwen3-VL-32B 8.7% smaller than NVFP4, at the same quality.

   Qwen3-VL-32B reads images and text. Tileward’s TWFP4 build is 19.96 GB against 21.86 GB for Red Hat’s NVFP4 build, and level with it on MMLU, MMStar, made-up facts and tool calls. Red Hat’s build stays closer to the original on next-word agreement.

   Model compressionMMLUQuality
   Experiment

   19.96 GB

   of weights, against 21.86 GB for NVFP4

   82.9%

   on MMLU, against 82.5% for NVFP4

   1.76 GiB

   less of the card for its weights than NVFP4

   Take a look →
4. R-009
   16 October 2026 · Benchmark

   We shrank Whisper large-v3 by 68%, below the size of turbo.

   Whisper large-v3 is a 3.09 GB speech recogniser. Tileward’s TWX4 build is 0.98 GB, smaller than the published turbo version. On nine of ten test sets it adds at most 0.3 points to the original’s error rate, and on Danish it adds 1.06.

   Model compressionQuality
   Experiment

   0.98 GB

   of weights, against 3.09 GB as published

   0.3 points

   at most, of error added on nine of ten sets; 1.06 on Danish

   152

   live calls on one 20 GB card, against 140

   Take a look →
5. R-008
   10 October 2026 · Benchmark

   We compressed VoiceChat by 74% and made it run in real time on one card.

   NVIDIA’s VoiceChat is a speech model that can call tools while it talks. Tileward’s TWX4 build is an 11.5 GB file against 44.4 GB. It picked the right tool on 55 of 80 spoken requests, level with the original’s 52, and chooses the original’s next word 95% of the time. With Tileward’s serving changes VoiceChat runs in real time on one card. The build runs from unpacked weights, so its speed and memory there match the original’s.

   Model compressionQuality
   Experiment

   11.5 GB

   of file, against 44.4 GB as published

   55 of 80

   right tool on spoken requests, against 52 for the original

   72 ms

   for each 80 ms frame with our serving changes, against 114 ms for NVIDIA’s open code

   Take a look →
6. R-007
   8 October 2026 · Benchmark

   We shrank PersonaPlex by 62% to run at 20ms a frame.

   NVIDIA’s PersonaPlex speech model is 16.7 GB as published. Tileward’s TWX4 build is 6.3 GB, needs 20 ms for each 80 ms frame against about 40 on one laptop, and picks the original’s next word 96% of the time. A standard 8-bit conversion stays closer, at 8.2 GB.

   Model compressionQuality
   Experiment

   6.3 GB

   of weights, against 16.7 GB as published

   20 ms

   of compute for each 80 ms frame, against about 40

   96%

   of the time, the same next word as the original

   Take a look →
7. R-006
   6 September 2026 · Benchmark

   Keep most of the answer rate on 95% less context (19×).

   From a 1,986-question LOCOMO set, Tileward Context answered 55% of answerable questions with 1,499 tokens each; full transcripts answered 63–67% with 28,365. A separate model judged every answer, and on unanswerable questions Context made up an answer far less often.

   Conversation memoryLOCOMOToken costQuality
   Experiment

   769 / 41

   questions won and lost against keeping the newest turns

   95%

   less context than sending everything (18.9×)

   55% vs 63–67%

   questions answered, against sending the whole transcript

   Take a look →
8. R-005
   4 September 2026 · Benchmark

   Give agents useful context, not the whole backlog.

   Across 104 AgentBench tasks, Tileward Context solved 37 with 523,028 context tokens; full history solved 24 with 1,651,372. The 68% (3.2×) reduction is stable, but 14–19 tasks can flip between identical runs, so the repeated ordering matters more than a fixed 13-task margin.

   Agent workflowsAgentBenchToken costQuality
   Experiment

   37 / 24

   tasks solved with Context / full history, of 104

   68%

   less context than sending everything (3.2×)

   14–19

   tasks the same setup flips against itself

   Take a look →
9. R-004
   31 August 2026 · Benchmark

   Improve recall without losing correct answers.

   On 140 scored LongMemEval-S questions, one recall upgrade moved the same store and model from 67 correct to 81: 14 wrong-to-right flips and no reversals. This is one self-judged run, and the upgrade itself remains undisclosed.

   Conversation memoryLongMemEvalQuality
   Experiment

   +10.0

   points on LongMemEval-S, paired

   14 / 0

   questions fixed / broken

   9→16

   LongMemEval-S temporal-reasoning, of 40

   Take a look →
10. R-003
    28 August 2026 · Benchmark

    A smaller model should not become a different model.

    Across two model families, compression did not create new drift in sensitive record-copying behavior. Model selection mattered far more, while a paired knowledge test measured the actual trade-off at 0.40 MMLU points on a 27B research build.

    Model compressionMMLUQuality
    Experiment

    26 of 30

    starved tokens shared across families

    0.0

    embedding difference after compression

    −0.40

    MMLU points, 27B research build against its original, paired p = 0.021

    Take a look →
11. R-002
    28 August 2026 · Governance

    Put policy in front of the model, not inside the prompt.

    A pre-inference gate turns a policy from advice to an enforceable, auditable decision. We tested it with 2,251 requests written to evade the rule and 50 benign controls. The payoff is a boundary your team can test before depending on it.

    GovernanceQuality
    Experiment

    2,251

    adversarial cases

    50

    crossed the gate

    5 of 50

    benign controls refused

    Take a look →
12. R-001
    25 August 2026 · Benchmark

    Stop paying every agent to reread everything.

    Across six multi-agent problems, Tileward Context cut billable tokens by 72.1% against broadcast context, including ingest, and produced the strongest code in the run. Wide research fan-out exposed a retrieval boundary, so test that shape first.

    Agent workflowsToken costQuality
    Experiment

    ≈ 72.1%

    reduction in billable tokens

    0.968

    coding, best of four

    5.33

    research panel, last of four

    Take a look →

No research write-ups match that filter yet.