Parallax 8B lambda: training checkpoints

These are research checkpoints from lambda, a training run of Parallax. Training was stopped at about 900B tokens, before the planned end of the run. The designated final checkpoint is step 73973 (900.8B tokens), marked in the table below. Exports saved later than the final checkpoint were saved after training stopped and are not the final checkpoint. Parallax is a mixture-of-experts language model trained on consumer GPUs spread over several countries. The hosts have no direct connections to each other and synchronize over the public internet.

• Base model. It is pretrained on a mixture of web, document, math, code, knowledge-focused and book text. It is not instruction-tuned, chat-tuned or safety-tuned, and it will continue text rather than follow instructions.
• Training snapshots. Every export, the final one included, is a snapshot taken before the planned ~973B tokens. Quality between exports is not monotonic.
• Research artifact. It is published so the training can be followed and inspected. It is not intended for production use.

Live training dashboard: parallax.chutes.ai.

Model

Training

• Data: seven stationary sources totalling about 1.09T tokens: two general pretraining mixtures of web, document, math and code text, an additional web set, two math sets, a knowledge-focused set and a books set. At about 899.6B tokens the data switched to two further sources while the learning-rate anneal that began at 877.28B continued. Training stopped at about 900B tokens, shortly after this switch. The planned training total was about 972.9B tokens.
• Optimization: the dense trunk is synchronized with a decoupled DiLoCo scheme over the internet. Routed experts are trained through low-rank adapters on the GPUs that use them, and the adapter updates are folded into full-precision masters that publish new ternary versions. The tech report describes the details.
• Learning rate: linear warm-up from 0 to 6e-4 over the first 19.66B tokens, then 6e-4 until 400B tokens, 3e-4 until 700B and 1.5e-4 until 877.28B. A live anneal started at 877.28B tokens on 2026-10-09, with multiplier m = 1 - 0.9 sqrt((T - 877.28B) / (972.90B - 877.28B)) and learning rate m times 1.5e-4. When training stopped, the multiplier had reached 0.5489 (learning rate 8.234e-5). The anneal was partial.
• Batch: 10 gradient-accumulation micro-batches per step at launch (about 8.5M tokens per fleet step with 26 machines), switched to 20 at about 100B tokens on 2026-10-04 at 22:08 UTC. This gives about 17M tokens per step with all 26 machines; tokens per step vary with fleet size.
• Fleet: 26 machines and 208 GPUs at launch. Seven machines were lost over 5-8 October, with roles reassigned automatically while training continued. At the end of the run, 19 machines were assigned, totalling 152 GPUs.
• Document masking: In training, attention and recurrent state are reset at document boundaries (up to 64 documents per 4096-token sequence; further boundaries are merged into the 64th document).
• Output logits: the learnable output logit scale is capped at 3.0 in training and inference (see Files and format).

Exports

One folder per export under exports/, named by training tokens (rounded down to whole billions) and fleet step. Exports were added about every 30 minutes while training ran. Training has stopped; the final checkpoint is marked in the table. Published exports are never modified or removed.

• This table lists the exports only. Final benchmarks of the final checkpoint, in percent, from the same lm-eval harness used for every model in the comparison: 35.4 on MMLU (cloze, 5-shot), 46.5 on ARC-Challenge (5-shot), 59.8 on HellaSwag (0-shot), and, on the full generative test splits, 14.3 on GSM8K (5-shot), 15.4 on MBPP (3-shot) and 12.2 on HumanEval (pass@1). No statistically significant aggregate difference was detected among the three late exports (paired 95% confidence intervals). The tech report has the full comparison, including other public base models.

Final checkpoint

exports/900B-tokens_step73973: 900.8B training tokens, step 73973.

hf download chutesai/parallax-8b-lambda --include "exports/900B-tokens_step73973/*" --local-dir parallax-8b-lambda
cd parallax-8b-lambda/exports/900B-tokens_step73973
tar -xf packed_experts.tar
sha256sum -c --quiet SHA256SUMS   # every file of the original export, byte for byte

Files and format

The files are in Parallax's native compact export format (inference only, no optimizer state). They are byte-identical to the export the training system produced:

The only change from the original export is packaging. The 4096 expert files are stored in one uncompressed tar to keep the repository's file count manageable. Extract it and check SHA256SUMS as shown above. Every upload was checked against the training system's own digests before and after it was published.

Logit-scale bound. The model's output logit scale is bounded: the forward pass uses exp(min(logit_scale_log, log 3.0)). The stored trunk tensor is the raw training parameter (it can sit slightly above the bound, e.g. from bf16 rounding), and each export records the bound in manifest.json (logit_scale: max, raw, effective) and in coverage.json (inference_policy.logit_scale_max). A loader must apply the recorded bound; the files themselves are left byte-identical.

Running it

This is a base model only. It is not chat- or instruction-tuned, so it does plain text completion: give it the start of a text and it continues it. It will not follow instructions or hold a conversation.

Standard transformers cannot load this format. Use the parallax-lambda branch of our llama.cpp fork, https://github.com/chutesai/llama.cpp. It adds a dedicated runtime, llama-parallax, for CPU (x86-64, ARM64) and Apple GPUs (Metal). Lambda's recurrent mixers (EDA) are implemented only on this branch. The fork's master branch runs the earlier kappa checkpoints and cannot run lambda.

A vLLM serving path for GPU servers is in preparation.

The finisher continuation

After lambda stopped, its final checkpoint (exports/900B-tokens_step73973, 900.8B tokens) was continued on a single 8-GPU server for a further 66.8B tokens, without the fleet. The mid-training mixture replays part of lambda's data and adds higher-quality web and reference text, books and other creative writing, scientific papers, code, math, instruction-style and question-answering text, synthetic long-document retrieval tasks and reasoning traces. The training context was raised in stages from 4,096 to 8,192, 16,384, 32,768 and 65,536 tokens (23.8B tokens over the first three lengths with dense global attention; then 6B at 16,384 tokens, during which, after about 1B tokens, the sparse global-attention selection was switched on; then 10B at 32,768 and 12B at 65,536 tokens with it on), followed by a closing 15B-token learning-rate anneal over a mix of 4,096- to 65,536-token sequences. The longest trained context is 65,536 tokens; a planned 131,072-token stage was skipped. The architecture, tokenizer and expert format are unchanged; the RoPE base of the global-attention layers was raised from 1e6 to 1.67e7 for the longer contexts. The finisher's final export is exports/finisher-967B-tokens_step31888 (967.6B tokens in total, finisher step 31888); like the lambda exports it is a base model with no instruction, chat or safety tuning. The Model and Training sections above describe the original lambda run.

Benchmarks of the finisher's final export, in percent, from the same lm-eval harness as the lambda numbers above; the lambda column repeats the final checkpoint's earlier scores on that harness for comparison:

Long context, finisher final only: needle-in-a-haystack retrieval 100, 100 and 98 at 16k, 32k and 64k tokens, and 63 at 128k (past the trained context), in percent over 100 items each; the RULER seven-task mean was 79.3, 72.4, 60.5, 49.7, 33.5 and 26.3 at 2k, 4k, 8k, 16k, 32k and 64k. The needle test uses the finisher's own synthetic long-context format, so it is in-distribution and measures retrieval the model was trained for; RULER is the external benchmark. These long-context numbers were measured with the training code with sparse global attention, not with the GGUF runtime below.

Ready-made GGUF

parallax-8b-lambda-finisher-967B-tokens.gguf at the repository root (2,133,006,592 bytes, 2.13 GB, sha256 4a3ed3cf2d4ca959a7a93f0dac9e4cd52b45e39355e4aa63244080b56916cd74, also in the .sha256 file next to it) was converted from exports/finisher-967B-tokens_step31888, the finisher's final export (967.6B training tokens in total, finisher step 31888). It keeps the export's bf16 trunk and stores the routed experts' exact ternary codes and per-row scales; no weight is requantized. Its global-attention layers run with sparse block selection and the finisher's RoPE base, as trained; the lambda GGUFs below run dense global attention. Each GGUF is a snapshot of one export. To use another export, convert it yourself (see tools/parallax/README.md on the branch).

The earlier snapshots, converted from lambda exports, remain at the repository root:

hf download chutesai/parallax-8b-lambda parallax-8b-lambda-finisher-967B-tokens.gguf --local-dir .
hf download NousResearch/Meta-Llama-3-8B tokenizer.json --local-dir .   # Llama 3 tokenizer

Build and run

Requires CMake, a C++17 compiler, and Python 3 with tokenizers for the text front end.

git clone -b parallax-lambda https://github.com/chutesai/llama.cpp && cd llama.cpp
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=ON   # -DGGML_METAL=OFF for CPU only
cmake --build build --target llama-parallax -j 8
build/bin/llama-parallax --backend metal --self-test               # or --backend cpu
pip install tokenizers

# Apple GPU
python tools/parallax/run.py --binary build/bin/llama-parallax \
  --model ../parallax-8b-lambda-finisher-967B-tokens.gguf --tokenizer ../tokenizer.json \
  --backend metal --experts lut9 --prompt 'The capital of France is' --predict 64

# CPU (x86-64 or ARM64)
python tools/parallax/run.py --binary build/bin/llama-parallax \
  --model ../parallax-8b-lambda-finisher-967B-tokens.gguf --tokenizer ../tokenizer.json \
  --backend cpu --experts lut8 --threads 8 --prompt 'The capital of France is' --predict 64

Generation is greedy by default. build/bin/llama-parallax --help lists sampling, context size (-c, default 4096), prompt scoring and benchmark options.

Speed and memory

Measured on an Apple M5 Max on AC power (batch 1, 128 generated tokens) with the bf16-trunk GGUF of the earlier 270B-tokens_step23035 export, which has the same architecture, using the same runtime code; speed does not depend on the weight values. The finisher GGUF additionally runs the sparse-attention indexer and block selection in its global-attention layers; it was not re-timed. Prompt times used --prefill-batch 128:

Peak resident memory was about 2.8 GB on Metal and 3.4 GB on CPU at the default 4,096-token context. Other Apple and ARM hardware will differ. ARM64 CPUs use NEON kernels. x86-64 CPUs use portable C++ that gives the same results but is not tuned for speed: on a shared, loaded 64-core AMD EPYC 9534 server it generated about 22 tokens/s with 16 threads.

Fidelity

The finisher GGUF was checked on CPU with lut against the training code's own model loaded from the same export, with FP32 activations in both, on the same 2,555 held-out positions from public text and code passages as the earlier releases, with sparse global attention in both. Overall mean KL was 1.7e-9 nats/token (per passage 1.5e-9 to 1.8e-9); the top-1 next token agreed at all 2,555 positions; the largest logit difference at any position was 0.0010; and three 48-token greedy continuations matched exactly. No passage showed the divergence seen with the lambda GGUFs, where a router near-tie made the two implementations pick different experts from one token onward.

The training code itself runs in bf16; against that bf16 forward the top-1 agreement was 94.9% (mean KL 0.0027, measured on the earlier 270B-tokens_step23035 export), which reflects the bf16 rounding of the reference, not an error of the runtime.

On the finisher GGUF, the largest per-layer relative RMS difference between Metal and CPU was about 2.7e-5 (logits 6.8e-6) at the final position of a 512-token text, and the two gave identical 48-token greedy continuations on three prompts. lut8 (CPU) quantizes expert lookup tables to 8 bits and is slightly approximate; lut and lut9 are exact.

Limitations of the runtime

• One sequence per process, greedy or simple sampled generation. No chat template, batching or server.
• Not integrated into llama-cli, llama-server or the standard llama.cpp model loader.
• Context is limited by -c. Inference uses a 2,048-token window for all sliding-window layers. With the finisher GGUF the global-attention layers attend to up to 16 selected blocks of 128 tokens per query, always including the current block (so inputs up to 2,048 tokens are attended in full); with the lambda GGUFs they are dense. The finisher was trained on sequences up to 65,536 tokens and lambda on 4,096, but the runtime was checked only on 512-token passages and 48-token continuations; longer inputs run (raise -c) but were not qualified in the runtime, and the long-context numbers above come from the training code.

Tech report

The Parallax tech report: https://parallax.chutes.ai/tech-report.pdf. It is AI-generated from the project's measurements and logs, and it is a living document that changes as the run progresses. It describes lambda, including the end of training and the final benchmark comparison.

Limitations

This is an early base model. It can produce incorrect, biased or nonsensical output. It has no alignment or safety tuning, and it is small and far from converged.

License

MIT.

_{Table updated 2026-10-21 23:47 UTC.}