Code and full history: github.com/Adam90a/local-256k-reasoning. Code is MIT; report and measurements CC BY 4.0; see NOTICE.md for third-party source excerpts in the test prompts.

local-256k-reasoning

Which open-weight model gives the best reasoning with ≥256K tokens of reliable context on one 24 GB GPU plus 128 GB of RAM, and which engine should run it? A measured answer, from November 2026.

Also on Hugging Face: datasets/timmyGPT/local-256k-reasoning.

Hardware: RTX PRO 4000 Blackwell (24 GB, the only GPU used) · Threadripper 7970X · 128 GB DDR5-4800 · llama.cpp 29c720e.

Result

Qwen3.8-Flash-Next, run on the Strata NVFP4 engine with the NVFP4 GPTQ weights (routed experts in system RAM, int8 KV, 294,912-token window). llama.cpp with Unsloth's UD-Q4_K_XL GGUF also passes, about 3.5× slower.

• 256K qualification: PASS on both engines. 10 consecutive runs at ≥262,144 input tokens (counted with the chat template), no truncation or errors, 13/13 on every run: 7 retrieval/aggregation checks, 5 cross-context reasoning tasks, and a standing instruction.

At ~262K input tokens	llama.cpp (UD-Q4_K_XL)	Strata (NVFP4 GPTQ)
10-run qualification	13/13 ×10	13/13 ×10
Prompt reading	5.7 min (~765 tok/s)	1.5 min (~2,857 tok/s)
Generation	29.7 tok/s	102.6 tok/s
Short-prompt generation	33.7 tok/s	110.7 tok/s
Peak VRAM / engine RAM	19.4 GB / 91 GB	23.4 GB / 76 GB
Claude Code, 3-bug fix task	2.4–3.4 min	32 s

• Same prompts as Claude Opus 5.5: on three byte-identical 262K prompts, Qwen (both 4-bit builds), Opus 5.5 and MiniMax-M3 all scored 13/13. On this benchmark the local model matches Opus. The benchmark doesn't separate the top models, so this is not a claim of overall parity.
• 512K: 13/13 on 2/2 runs (llama.cpp, YaRN ×4, q8_0 KV). About 1M does not fit.
• AIME 2026 (15 problems): 14/15 (Opus 5.5: 15/15). Measured on llama.cpp.
• Claude Code / Hermes driving the local model: all bug-fix tasks solved, tests untouched.
• Weakest area: hard contest coding (LiveCodeBench v6 hard: 3/10 local vs 6/10 Opus).
• Caveat on Strata: a young, single-maintainer project (Windows-first; built from source here). AIME and coding were not re-run on it.

All candidates (llama.cpp unless noted)

Model (all on the same machine)	256K	512K	~1M	AIME 2026	LCB hard
Qwen3.8-Flash-Next Q4	PASS 10/10	13/13 ×2	doesn't fit	14/15	3/10
Qwen3.8-Flash-Next IQ3_S (3-bit)	13/13 ×3	–	–	13/15	–
Nemotron 3.5 Super NVFP4	11/12	10/13	runs, no usable answer	13/15	4/10
Nemotron 3.5 Bolt IQ2	10/12*	–	–	11/15	2/10
Claude Opus 5.5 (cloud reference)	13/13 ×3	–	–	15/15	6/10
MiniMax-M3 (cloud reference)	13/13 ×3	–	–	13/15	6/10

* needed a GGUF context-length override (experimental).

Full write-up: research/final-report.md and research/strata-trial.md. Also: 256K qualification · quantization study · architecture notes, including HGA · fine-tuning assessment · deployment.

Practical findings

• Size the window for input plus reasoning. A 278K window with a 262K prompt left too little room, and one run ran out of space mid-reasoning.
• Build llama.cpp against one CUDA version. CUDA 13 headers with the CUDA 12 runtime corrupt cudaDeviceProp, and every kernel needing dynamic shared memory then fails with "invalid argument". See research/environment.md.
• MTP speculative decoding gives +31% decode on short prompts but −15% at 262K. Keep it off for long context.
• Qwen3.8-Flash-Next's stock chat template rejects Claude Code's mid-conversation system messages. A one-line patched template is in configs/.
• Ollama's Vulkan backend ignores CUDA_VISIBLE_DEVICES. Pin it by disabling Vulkan.

Reproduce

# Strata (recommended): build sergqwer/strata-nvfp4 for your GPU, then
scripts/serve_strata.sh configs/strata-gptq-288k.json      # paths inside the JSON point at this machine; edit them
benchmarks/context/longctx_openai.py --url http://127.0.0.1:8091 --tokens 262144 --seed 101 --out strata.json
# llama.cpp
scripts/serve.sh configs/qwen38fn-q4-yarn2-288k.env        # edit the GPU UUID in serve.sh for your card
benchmarks/context/longctx.py --tokens 262144 --seed 101 --out out.json
benchmarks/context/longctx.py --prompt-file results/prompts/v2/262144_s21.json --tokens 262144 --out same_prompt.json
benchmarks/reasoning/quality.py --task aime --out aime.jsonl   # needs benchmarks/data (see its README)

Licensing: code MIT; report and measurements CC BY 4.0; test prompts contain third-party source excerpts, see NOTICE.md.

---

Repository layout (project notes)

Execution of ~/Documents/LLM-slow.md: find, test, and deploy the strongest local reasoning/coding model with ≥256K reliable context on this workstation, using only the RTX PRO 4000 Blackwell.

Layout

• research/ — environment, model research, triage (CSV + MD), architecture, hardware feasibility, budget.
• benchmarks/context/longctx.py — long-context qualification test (planted facts in real source code, exact token count via server tokenizer + chat template, automatic scoring).
• scripts/serve.sh <configs/x.env> — starts llama-server pinned to the PRO 4000 by UUID.
• scripts/run_ctx.sh <label> <tokens> <seed> — one test run with resource monitoring.
• scripts/monitor.py — 1 Hz sampler: VRAM, GPU util/power/temp/clock, RAM, swap, CPU, server RSS.
• results/raw/<label>/ — per-run JSON, logs, monitor CSV and peaks.

Engine

llama.cpp built from upstream commit 29c720e (2026-11-02) in ~/llama.cpp-lrr (CUDA 13.0, sm_120). The older ~/llama.cpp (June) is untouched.